Zbornik 20. mednarodne multikonference INFORMACIJSKA DRUŽBA - IS 2017 Zvezek A Proceedings of the 20th International Multiconference INFORMATION SOCIETY - IS 2017 Volume A Slovenska konferenca o umetni inteligenci Slovenian Conference on Artificial Intelligence Uredili / Edited by Matjaž Gams, Mitja Luštrek, Rok Piltaver http://is.ijs.si 9.–13. oktober 2017 / 9–13 October 2017 Ljubljana, Slovenia Zbornik 20. mednarodne multikonference INFORMACIJSKA DRUŽBA – IS 2017 Zvezek A Proceedings of the 20th International Multiconference INFORMATION SOCIETY – IS 2017 Volume A Slovenska konferenca o umetni inteligenci Slovenian Conference on Artificial Intelligence Uredili / Edited by Mitja Luštrek, Rok Piltaver, Matjaž Gams http://is.ijs.si 9. - 13. oktober 2017 / 9th – 13th October 2017 Ljubljana, Slovenia Uredniki: Mitja Luštrek Odsek za inteligentne sisteme Institut »Jožef Stefan«, Ljubljana Rok Piltaver Celtra, d. o. o. in Odsek za inteligentne sisteme Institut »Jožef Stefan«, Ljubljana Matjaž Gams Odsek za inteligentne sisteme Institut »Jožef Stefan«, Ljubljana Založnik: Institut »Jožef Stefan«, Ljubljana Priprava zbornika: Mitja Lasič, Vesna Lasič, Lana Zemljak Oblikovanje naslovnice: Vesna Lasič Dostop do e-publikacije: http://library.ijs.si/Stacks/Proceedings/InformationSociety Ljubljana, oktober 2017 Kataložni zapis o publikaciji (CIP) pripravili v Narodni in univerzitetni knjižnici v Ljubljani COBISS.SI-ID=1537600707 ISBN 978-961-264-112-2 (pdf) PREDGOVOR MULTIKONFERENCI INFORMACIJSKA DRUŽBA 2017 Multikonferenca Informacijska družba (http://is.ijs.si) je z dvajseto zaporedno prireditvijo osrednji srednjeevropski dogodek na področju informacijske družbe, računalništva in informatike. Letošnja prireditev je ponovno na več lokacijah, osrednji dogodki pa so na Institutu »Jožef Stefan«. Informacijska družba, znanje in umetna inteligenca so spet na razpotju tako same zase kot glede vpliva na človeški razvoj. Se bo eksponentna rast elektronike po Moorovem zakonu nadaljevala ali stagnirala? Bo umetna inteligenca nadaljevala svoj neverjetni razvoj in premagovala ljudi na čedalje več področjih in s tem omogočila razcvet civilizacije, ali pa bo eksponentna rast prebivalstva zlasti v Afriki povzročila zadušitev rasti? Čedalje več pokazateljev kaže v oba ekstrema – da prehajamo v naslednje civilizacijsko obdobje, hkrati pa so planetarni konflikti sodobne družbe čedalje težje obvladljivi. Letos smo v multikonferenco povezali dvanajst odličnih neodvisnih konferenc. Predstavljenih bo okoli 200 predstavitev, povzetkov in referatov v okviru samostojnih konferenc in delavnic. Prireditev bodo spremljale okrogle mize in razprave ter posebni dogodki, kot je svečana podelitev nagrad. Izbrani prispevki bodo izšli tudi v posebni številki revije Informatica, ki se ponaša s 40-letno tradicijo odlične znanstvene revije. Odlične obletnice! Multikonferenco Informacijska družba 2017 sestavljajo naslednje samostojne konference:  Slovenska konferenca o umetni inteligenci  Soočanje z demografskimi izzivi  Kognitivna znanost  Sodelovanje, programska oprema in storitve v informacijski družbi  Izkopavanje znanja in podatkovna skladišča  Vzgoja in izobraževanje v informacijski družbi  Četrta študentska računalniška konferenca  Delavnica »EM-zdravje«  Peta mednarodna konferenca kognitonike  Mednarodna konferenca za prenos tehnologij - ITTC  Delavnica »AS-IT-IC«  Robotika Soorganizatorji in podporniki konference so različne raziskovalne institucije in združenja, med njimi tudi ACM Slovenija, SLAIS, DKZ in druga slovenska nacionalna akademija, Inženirska akademija Slovenije (IAS). V imenu organizatorjev konference se zahvaljujemo združenjem in inštitucijam, še posebej pa udeležencem za njihove dragocene prispevke in priložnost, da z nami delijo svoje izkušnje o informacijski družbi. Zahvaljujemo se tudi recenzentom za njihovo pomoč pri recenziranju. V 2017 bomo petič podelili nagrado za življenjske dosežke v čast Donalda Michija in Alana Turinga. Nagrado Michie-Turing za izjemen življenjski prispevek k razvoju in promociji informacijske družbe bo prejel prof. dr. Marjan Krisper. Priznanje za dosežek leta bo pripadlo prof. dr. Andreju Brodniku. Že šestič podeljujemo nagradi »informacijska limona« in »informacijska jagoda« za najbolj (ne)uspešne poteze v zvezi z informacijsko družbo. Limono je dobilo padanje slovenskih sredstev za akademsko znanost, tako da smo sedaj tretji najslabši po tem kriteriju v Evropi, jagodo pa »e-recept«. Čestitke nagrajencem! Bojan Orel, predsednik programskega odbora Matjaž Gams, predsednik organizacijskega odbora i FOREWORD - INFORMATION SOCIETY 2017 In its 20th year, the Information Society Multiconference (http://is.ijs.si) remains one of the leading conferences in Central Europe devoted to information society, computer science and informatics. In 2017 it is organized at various locations, with the main events at the Jožef Stefan Institute. The pace of progress of information society, knowledge and artificial intelligence is speeding up, and it seems we are again at a turning point. Will the progress of electronics continue according to the Moore’s law or will it start stagnating? Will AI continue to outperform humans at more and more activities and in this way enable the predicted unseen human progress, or will the growth of human population in particular in Africa cause global decline? Both extremes seem more and more likely – fantastic human progress and planetary decline caused by humans destroying our environment and each other. The Multiconference is running in parallel sessions with 200 presentations of scientific papers at twelve conferences, round tables, workshops and award ceremonies. Selected papers will be published in the Informatica journal, which has 40 years of tradition of excellent research publication. These are remarkable achievements. The Information Society 2017 Multiconference consists of the following conferences:  Slovenian Conference on Artificial Intelligence  Facing Demographic Challenges  Cognitive Science  Collaboration, Software and Services in Information Society  Data Mining and Data Warehouses  Education in Information Society  4th Student Computer Science Research Conference  Workshop Electronic and Mobile Health  5th International Conference on Cognitonics  International Conference of Transfer of Technologies - ITTC  Workshop »AC-IT-IC«  Robotics The Multiconference is co-organized and supported by several major research institutions and societies, among them ACM Slovenia, i.e. the Slovenian chapter of the ACM, SLAIS, DKZ and the second national engineering academy, the Slovenian Engineering Academy. In the name of the conference organizers we thank all the societies and institutions, and particularly all the participants for their valuable contribution and their interest in this event, and the reviewers for their thorough reviews. For the fifth year, the award for life-long outstanding contributions will be delivered in memory of Donald Michie and Alan Turing. The Michie-Turing award will be given to Prof. Marjan Krisper for his life-long outstanding contribution to the development and promotion of information society in our country. In addition, an award for current achievements will be given to Prof. Andrej Brodnik. The information lemon goes to national funding of the academic science, which degrades Slovenia to the third worst position in Europe. The information strawberry is awarded for the medical e-recipe project. Congratulations! Bojan Orel, Programme Committee Chair Matjaž Gams, Organizing Committee Chair ii KONFERENČNI ODBORI CONFERENCE COMMITTEES International Programme Committee Organizing Committee Vladimir Bajic, South Africa Matjaž Gams, chair Heiner Benking, Germany Mitja Luštrek Se Woo Cheon, South Korea Lana Zemljak Howie Firth, UK Vesna Koricki Olga Fomichova, Russia Mitja Lasič Vladimir Fomichov, Russia Robert Blatnik Vesna Hljuz Dobric, Croatia Aleš Tavčar Alfred Inselberg, Israel Blaž Mahnič Jay Liebowitz, USA Jure Šorn Huan Liu, Singapore Mario Konecki Henz Martin, Germany Marcin Paprzycki, USA Karl Pribram, USA Claude Sammut, Australia Jiri Wiedermann, Czech Republic Xindong Wu, USA Yiming Ye, USA Ning Zhong, USA Wray Buntine, Australia Bezalel Gavish, USA Gal A. Kaminka, Israel Mike Bain, Australia Michela Milano, Italy Derong Liu, Chicago, USA Toby Walsh, Australia Programme Committee Bojan Orel, chair Mitja Luštrek Niko Schlamberger Franc Solina, co-chair Marko Grobelnik Stanko Strmčnik Viljan Mahnič, co-chair Nikola Guid Jurij Šilc Cene Bavec, co-chair Marjan Heričko Jurij Tasič Tomaž Kalin, co-chair Borka Jerman Blažič Džonova Denis Trček Jozsef Györkös, co-chair Gorazd Kandus Andrej Ule Tadej Bajd Urban Kordeš Tanja Urbančič Jaroslav Berce Marjan Krisper Boštjan Vilfan Mojca Bernik Andrej Kuščer Baldomir Zajc Marko Bohanec Jadran Lenarčič Blaž Zupan Ivan Bratko Borut Likar Boris Žemva Andrej Brodnik Janez Malačič Leon Žlajpah Dušan Caf Olga Markič Saša Divjak Dunja Mladenič Tomaž Erjavec Franc Novak Bogdan Filipič Vladislav Rajkovič Andrej Gams Grega Repovš Matjaž Gams Ivan Rozman iii Invited lecture AN UPDATE FROM THE AI & MUSIC FRONT Gerhard Widmer Institute for Computational Perception Johannes Kepler University Linz (JKU), and Austrian Research Institute for Artificial Intelligence (OFAI), Vienna Abstract Much of current research in Artificial Intelligence and Music, and particularly in the field of Music Information Retrieval (MIR), focuses on algorithms that interpret musical signals and recognize musically relevant objects and patterns at various levels -- from notes to beats and rhythm, to melodic and harmonic patterns and higher-level segment structure --, with the goal of supporting novel applications in the digital music world. This presentation will give the audience a glimpse of what musically "intelligent" systems can currently do with music, and what this is good for. However, we will also find that while some of these capabilities are quite impressive, they are still far from (and do not require) a deeper "understanding" of music. An ongoing project will be presented that aims to take AI & music research a bit closer to the "essence" of music, going beyond surface features and focusing on the expressive aspects of music, and how these are communicated in music. This raises a number of new research challenges for the field of AI and Music (discussed in much more detail in [Widmer, 2016]). As a first step, we will look at recent work on computational models of expressive music performance, and will show some examples of the state of the art (including the result of a recent musical 'Turing test'). References Widmer, G. (2016). Getting Closer to the Essence of Music: The Con Espressione Manifesto. ACM Transactions on Intelligent Systems and Technology 8(2), Article 19. iv KAZALO / TABLE OF CONTENTS Slovenska konferenca o umetni inteligenci / Slovenian Conference on Artificial Intelligence .......................... 1 PREDGOVOR / FOREWORD ................................................................................................................................. 3 PROGRAMSKI ODBORI / PROGRAMME COMMITTEES ..................................................................................... 5 Artificial Intelligence in 2017 / Gams Matjaž ........................................................................................................... 7 Comparison of feature ranking approaches for discovery of rare genetic variants related to multiple sclerosis / Petković Matej, Vidmar Lovro, Tanevski Jovan , Peterlin Borut , Maver Aleš , Džeroski Sašo.................................................................................................................................................................. 11 Modelling of dynamical systems: a survey of tools and a case study / Peev Gjorgi, Simidjievski Nikola, Džeroski Sašo .................................................................................................................................................. 15 Vpliv različnega prenosnega kanala pri referenčnih in testnih posnetkih na forenzično verifikacijo govorcev / Šef Tomaž, Blatnik Robert ............................................................................................................. 19 JSI Sound – platforma za enostavno klasifikacijo zvočnih posnetkov: Demonstracija na zvokih živali / Budna Borut, Gjoreski Martin, Gradišek Anton, Gams Matjaž ......................................................................... 23 Bat classification using Deep Neural Network / Bizjak Jani, Gradišek Anton, Stepančič Luka, Presetnik Primož............................................................................................................................................................... 27 Globoke nevronske mreže in matrična faktorizacija / Petelin Gašper, Kononenko Igor ...................................... 31 Optimiranje časa in porabe goriva v modelih človeške vožnje / Dovgan Erik, Bratko Ivan, Sodnik Jaka, Filipič Bogdan ................................................................................................................................................... 35 User-friendly multi-objective learning of accurate and comprehensible hybrid-trees / Novak Benjamin, Piltaver Rok, Gams Matjaž ............................................................................................................................... 39 Automatic Tennis Analysis with the use of Machine Learning and Multi-Objective Optimization / Mlakar Miha, Luštrek Mitja ........................................................................................................................................... 43 Anytime Benchmarking of Budget-Dependent Algorithmswith the COCO Platform / Tušar Tea, Hansen Nikolaus, Brockhoff Dimo ................................................................................................................................. 47 Criteria for Co-existence of GM and Conventional Maize Production / Debeljak Marko, Leprince Florence, Džeroski Sašo, Trajanov Aneta ........................................................................................................ 51 Knowledge Discovery from Complex Ecological Data: Exploring Syrphidae Species in Agricultural Landscapes / Debeljak Marko, Kuzmanovski Vladimir, Tosser Veronique, Trajanov Aneta .......................... 55 A Comparison of DEXi, AHP and ANP Decision Models for Evaluation of Tourist Farms / Dergan Tanja ......... 59 A State-Transition Decision Support Model for Medication Change of Parkinson's Disease Patients / Boshkoska Biljana Mileva, Gatsios Dimitris, Kostas M. Tsiouris, Miljković Dragana, Rigas George, Fotiadis Dimitrios, Valmarska Anita, Konitsiotis Spiros, Bohanec Marko ........................................................ 63 Designing a Personal Decision Support System for Congestive Heart Failure Management / Bohanec Marko, Dovgan Erik, Maslov Pavel, Vodopija Aljoša, Luštrek Mitja, Puddu Paolo Emilio, Schiariti Michele, Ciancarelli Maria Costanza, Baert Anneleen, Pardaens Sofie, Clays Els ......................................... 67 Continuous Blood Pressure Estimation from PPG Signal / Slapničar Gašper, Marinko Matej, Luštrek Mitja ...... 71 Recognizing Hand-Specific Activities with a Smartwatch Placed on Dominant or Non-dominant Wrist / Cvetković Božidara, Drobnič Vid, Luštrek Mitja ............................................................................................... 75 R-R vs GSR – An inter-domain study for arousal recognition / Gjoreski Martin, Mitrevski Blagoj, Luštrek Mitja, Gams Matjaž ........................................................................................................................................... 79 Predicting office’s ambient parameters / Janko Vito, Stensbye Andreas R., Luštrek Mitja .................................. 83 Real-time content optimization in digital advertisement / Vodopivec Tom, Sušelj Gregor, Sluga Davor, Piltaver Rok, Ilc Nejc, Košir Domen ................................................................................................................. 87 Indeks avtorjev / Author index ................................................................................................................................ 91 v vi Zbornik 20. mednarodne multikonference INFORMACIJSKA DRUŽBA – IS 2017 Zvezek A Proceedings of the 20th International Multiconference INFORMATION SOCIETY – IS 2017 Volume A Slovenska konferenca o umetni inteligenci Slovenian Conference on Artificial Intelligence Uredili / Edited by Mitja Luštrek, Rok Piltaver, Matjaž Gams http://is.ijs.si 12. - 13. oktober 2017 / 12th – 13th October 2017 Ljubljana, Slovenia 1 2 PREDGOVOR V letu 2017 smo bili spet priča neverjetnim dosežkom umetne inteligence, ki na čedalje več področjih prekaša človeške sposobnosti. Velja omeniti poker Texas hold'em brez omejitev pri višini stav (ki je precej bolj kompleksen od že rešene različice z omejitvami) in strateško računalniško igro Dota 2, kjer so se do sedaj ljudje uspešno upirali programom umetne inteligence, sedaj pa je v igri ena na ena umetna inteligenca pokazala premoč. Podobno dobro rešuje tudi resnejše probleme, npr. prepoznavanje rakavih tkiv za zgodnjo diagnozo, kjer pa opažamo počasen prenos dosežkov iz raziskovalnih laboratorijev v prakso. Umetna inteligenca že sedaj ljudem pomaga na veliko področjih in celo rešuje življenja. Trendi kažejo, da bo naslednje leto še bolj koristna in prijazna. In naslednja leta še bolj. Mnoge zanimive dosežke umetne inteligence lahko spoznamo tudi na Slovenski konferenci o umetni inteligenci (SKUI). Letos smo sprejeli 21 prispevkov, kar so trije več kot lani. Kot pretekla leta jih je največ z Instituta »Jožef Stefan«. Obžalujemo, da jih je manj kot lani prispevala Fakultete za računalništvo in informatiko, ki ima skupaj z Institutom vodilno vlogo pri raziskavah umetne inteligence v Sloveniji, pozdravljamo pa dva zelo kakovostna prispevka iz industrije. Upamo, da bo prispevkov iz industrije in nasploh izven Instituta prihodnja leta še več, saj je ključen cilj SKUI povezovanje vseh slovenskih raziskovalcev umetne inteligence, čeprav na konferenci niso nič manj dobrodošli tudi prispevki iz drugih držav. SKUI je naslednica konference Inteligentni sistemi, ki je sestavni del multikonference Informacijska družba že od njenega začetka leta 1997. Letos tako skupaj s celotno multikonferenco praznuje 20. obletnico. Ker poleg tega Slovensko društvo za umetno inteligenco (SLAIS) – ki SKUI šteje za svojo konferenco – praznuje 25. obletnico, smo se odločili razširjene različice najboljših prispevkov povabiti v posebno številko revije Informatica o umetni inteligenci. Objava najboljših prispevkov z Informacijske družbe v Informatici je že dolga tradicija, ki pa jo bomo letos s posebno številko revije, kjer bomo objavili izbrane raziskave umetne inteligence v Sloveniji, še oplemenitili. Mitja Luštrek, Rok Piltaver, Matjaž Gams 3 FOREWORD 2017 has brought many exciting achievements of artificial intelligence, which is proving superior to humans in increasingly many fields. Two examples are no-limit Texas hold’em poker (which is substantially more complex than the already solved limit version) and the computer strategy game Dota 2. In both cases, artificial intelligence has not been able to match the best humans so far, but this changed this year. Artificial intelligence is also solving more serious problems, such as the identification of cancerous tissue to enable early diagnosis; unfortunately, though, such achievements are not translated from research laboratories to practice as quickly as we may wish. However, artificial intelligence is already helping people in many fields and even saving lives. Trends indicate that it will be even more useful and friendly next year, and more so the years after that. Slovenian Conference on Artificial Intelligence (SCAI) is a venue where one can learn about many achievements of artificial intelligence. 21 papers were accepted this year, which is three more than previous year. As in past years, most of them were from Jožef Stefan Institute. We regret that the Faculty of Computer and Information Science, which shares the leading role in artificial intelligence research in Slovenia with the Institute, contributed fewer papers this year; however, we are glad to have received two very high-quality papers from the industry. We hope for even more papers from the industry and other institutions outside the Institute in the following years, since a key objective of the conference is bringing together all Slovenian artificial intelligence researchers, although international papers are of course equally welcome. SCAI is the successor of the Intelligent Systems conference, which has been a part of the Information Society multiconference since its establishment in 1997. The conference – together with the whole multiconference – thus celebrates its 20th anniversary this year. In addition, Slovenian Artificial Intelligence Society (SLAIS), which is the main supporter of SCAI, celebrates its 25th anniversary. Because of that, the extended versions of the best papers will be invited to a special issue of the Informatica journal on artificial intelligence. Publishing the best papers from the Information Society conference in the Informatica journal has a long tradition, but this year the best SCAI papers will find themselves in the company of other selected papers on the Slovenian research on artificial intelligence. Mitja Luštrek, Rok Piltaver, Matjaž Gams 4 PROGRAMSKI ODBOR / PROGRAMME COMMITTEE Mitja Luštrek, IJS (co-chair) Rok Piltaver, IJS (co-chair) Matjaž Gams, IJS (co-chair) Marko Bohanec Tomaž Banovec Cene Bavec Jaro Berce Marko Bonač Ivan Bratko Dušan Caf Bojan Cestnik Aleš Dobnikar Bogdan Filipič Nikola Guid Borka Jerman Blažič Tomaž Kalin Marjan Krisper Marjan Mernik Vladislav Rajkovič Ivo Rozman Niko Schlamberger Tomaž Seljak Miha Smolnikar Peter Stanovnik Damjan Strnad Peter Tancig Pavle Trdan Iztok Valenčič Vasja Vehovar Martin Žnidaršič 5 6 Artificial Intelligence in 2017 Matjaž Gams Jozef Stefan Institute Jamova 39, 1000 Ljubljana, Slovenia matjaz.gams@ijs.si ABSTRACT 2. IJCAI 2017 In recent years, AI is facing incredibly fast progress. In this paper we review a couple of major new AI-related achievements and The 26th International Joint Conference on Artificial Intelligence events. Among them, IJCAI 2017 as the cover AI worldwide was held in Melbourne, Australia in August 2017 [6]. Melbourne conference presented major scientific and industrial achievements is world's most liveable city for seventh year running and indeed it along with several discussions and panels. Among them was AI is safe, clean, not crowded, full of green nature and architectural superiority in the unlimited Texas hold’em poker and Dota 2. wonders. Both games were slightly limited, e.g. only 2 players instead of 10 in Dota 2, but the games itself included all major components such as bluffing with hidden cards or properties of dynamic strategic game with global and local decision making. Deep Neural Networks continue their excellence in visual recognition tasks and in real-life diagnostics, such as diagnosing which tissue contains malignant cancer cells, exceed best human experts in more and more diseases. Among broader influence of AI on human future life, the ban of autonomous weapons was steadily promoted and as a result, the asilomar principles were defined for the first time. The principles present an attempt to provide guidelines for human-beneficial AI, the one that would prevent possibilities for AI to turn into human- harmful ways. The aim of the paper is to bring these issues to our Figure 1: The growth of IJCAI papers in recent years. society through presentation and discussions. Keywords The AI growth is indicated by the number of papers submitted to Artificial intelligence, AI principles, Future of life institute the IJCAI conference (Figure 1). In 2016 in New York there were 2.294 papers submitted while in 2017 in Melbourne, 2540 papers were reviewed. The growth was steady from 2009 on. 1. INTRODUCTION The progress of artificial intelligence (AI) is fast and often surprisingly efficient even for AI professionals [5]. Each year there are scores of new achievements in academia, gaming, industry, and real life. There are also practical modifications of the way we live and work. For example, autonomous vehicles are improving constantly, they are introduced into more and more countries. In Europe, the goal to introduce similar legislation promoting the drones and autonomous vehicles alike by the EU Commissioner Violeta Bulc has not been successful yet, while several EU countries have modified their traffic laws accordingly and USA has recently changed its legislation to promote faster implementation of autonomous vehicle into real life. In Slovenia, Figure 2: Papers per countries at IJCAI 2017. where the Justice Minister Goran Klemenčič is intensively trying to modernize the legal system despite the resistance of mainly Study of papers submitted per country (Figure 2) at IJCAI 2017 status-quo majority, the drones are prohibited to spam the space, indicates that the majority of them was from China (37%), second but as it is becoming a European habit, the bureaucratic viewpoint EU (18%) and third US (18%). prohibits the use of drones and autonomous vehicles also for On September 1, Vladimir Putin speaking with students warned scientific purposes. As a result, Slovenian researchers are that whoever cracks artificial intelligence will 'rule the world' [9]. developing drones and autonomous cars illegally, but luckily Will that be China since it already submits the major bulk of AI nobody charges them for that. This is just one example how the papers? Or will it be USA since most of the awards were given to political and legal system is lagging behind the progress of USA researchers? artificial intelligence and ICT – information and communication It is not only the number of AI papers from China, the industry technologies. achievements are astonishing as well. One might not be as familiar with the Chinese solutions as with Google or Amazon AI 7 systems, but Chinese systems are close to the top. For example, in 2017 China’s Alibaba Group Holding Ltd introduced a cut-price voice assistant speaker, similar to Amazon.com Inc’s “Echo”. It is named “Tmall Genie” and costs $73, significantly less than western counterparts by Amazon and Alphabet Inc’s Google, which range around $150. Similarly, Baidu, China’s top search engine, recently launched a device based on its own Siri-like “Duer OS” system. Alibaba and China’s top tech firms have ambitions to become world leaders in artificial intelligence as companies. Figure 3: Error of DNNs on ImageNet through years. In terms of overall several scientific and practical achievements There were several demonstrations and competitions at IJCAI presented at IJCAI 2017, two games stood out as another example 2017, including the traditional Angry birds competition. Most of AI beating the best human counterparts: unlimited Texas attractive, however, were soccer competitions with off-line Nao hold’em poker (10 on 160 possibilities) and Dota 2. Both games robots that were not trained or advised as a team, but performed were slightly limited - in poker, there are only two players, and on their own in a group proclaimed at the spot. Unfortunately, the Dota 2 was also reduced to only two players instead of 10. local computing powers are at the level of a mobile phone, Nevertheless, both games are most-played human games with insufficient for good play. In Figure 4 one can see robots award funds going into tens of millions. Both games are quite wondering around and searching for a ball. Still, they different from formal games like chess or Go. For example, poker demonstrated some quite cunning properties, e.g. precise kicking included all major components of human bluffing interactions and the ball into the goal under the desired angle compared to the foot. hidden cards. Dota 2 was constructed in a way that fast computers had no advantage and the outcome of a game was dependent on strategic plans with global and local decision making, and adapting to the adversary. From Wikipedia: “Dota 2 is originally played in matches between two teams of five players, with each team occupying and defending their own separate base on the map. Each of the ten players independently controls a powerful character, known as a "hero", who all have unique abilities and differing styles of play. During a match, the player collects experience points and items for their heroes in order to successfully fight the opposing team's heroes, who are doing the same. A team wins by being the first to destroy a large structure located in the opposing team's base, called the "Ancient", which is guarded by defensive towers.” Regarding the methods, reinforcement learning and deep neural Figure 4: Soccer competition of independent individual Nao networks were somehow most common applied, however, the AI robots, dynamically assembled into teams at IJCAI 2017. field was presented through over 10 major areas. Next year will be of particular interest. ICML with 3500 Deep Neural Networks continue their excellence in visual attendees, IJCAI+ECAI with 2500, AAMAS with 700, ICCBR recognition tasks and in real-life diagnostics, such as diagnosing with 250 and SOCS with 50 attendees will be hosted at which tissue contains malignant cancer cells, exceed best human Stockholm in a 2-week event, July 2018. Wishful jokes are experts in more and more diseases. There are several tasks, e.g. emerging that the critical mass of 6-7000 attendees will provide recognition of faces from a picture where DNNs recognized the critical mass to ignite the general intelligence or even hundreds of faces in seconds, a result no human can match. Figure superintelligence [2, 8, 10]. 3 demonstrates the progress of DNNs in visual tasks: around 2015 the visual recognition in specific domains was comparable to 3. BAN OF AUTONOMOUS WEAPONS humans. Now, it is surpassed humans quite significantly – again, in particular visual tests. Due to the vicinity of the Syrian conflict it is interesting to The effects of only visual superiority are astonishing on its own. observe the level of sophistication of ICT solutions. ISIL, despite For example, eye analyses enable detecting certain diseases like its technical inferiority, was the first to use slightly modified cancer or Alzheimer [3]. Furthermore, DNN studies of facial industrial drones to drop small bombs on the infantry. They also properties enable detecting sexual orientation, IQ, and political use remotely controlled weapons such as machineguns. However, orientation. When shown five photos of each man, a recent system the often used suicide industrial cars, fully loaded with explosives was able to correctly select the man's sexuality 91 per cent of the and shielded by attached armor plates, are still driven by time, while humans were able to perform the same task with less vulnerable humans and not by remote controls on both sides. than 70% accuracy [7]. This Stanford University study alone None of these weapons falls into the category of fully autonomous confirmed that homosexuality is very probably of genetic origin. weapons AI scientists propose to ban since they don’t decide on The consequences of one single study can be profound. Will job its own when to fire. applications be determined also by the DNN study of facial There are two major reasons for the proposed ban: properties? Will dictatorship countries prosecuting homosexuality punish their citizens on the basis of their faces?  The fully autonomous weapons will likely make the war inhumane whereas humans – if war cannot be avoided – need 8 some rule of engagement to preserve some level of humanity process generated improved versions of the principles. Finally, and prevent too extreme human suffering. they surveyed the full set of attendees to determine the level of  This is one of preconditions on the road to prevent support for each version of each principle. superintelligence to go viral, malignant [2, 8, 10]. After the consuming and meticulous process, a high level of consensus emerged around many of the statements during that There is some reason for celebrating the first successes of the pro- final survey. The final list retained the principles if at least 90% of ban efforts – the movement is spreading through the social media the attendees agreed on them. The 23 principles were grouped since it started years ago by scientists like Toby Walsh or Stuart into research strategies, data rights and future issues including Russel and is currently coordinated by Mary Wareham. Slovenia potential superintelligence, signed by those wishing to associate is involved at national level where 4 societies (SLAIS for artificial their name with the list. The principles will hopefully provide intelligence, DKZ for cognitive science, Informatica for some guidelines as to how the power of AI can be used to informatics, ACM Slovenia for computer science) assembled a improve everyone’s lives in coming years. letter and sent it to the UN and Slovenian government, while lately the Slovenian AI society SLAIS submitted a letter to the At the web page of the event on the web pages of the Future of European national communities to join activities in this direction. Life Institute [4], the following original presentations can be Our initiative was also debated at the EurAI meeting at IJCAI obtained with additional interviews on the consequent links 2017. Artificial intelligence has already provided beneficial tools that are used every day by people around the world. Its continued Second, Elon Mask and CEOs of 155 robotic companies development, guided by the following principles, will offer assembled a letter, in which they write “Once developed, lethal amazing opportunities to help and empower people in the decades autonomous weapons will permit armed conflict to be fought at a and centuries ahead. scale greater than ever, and at timescales faster than humans can comprehend. These can be weapons of terror, weapons that 4.1 Research Issues despots and terrorists use against innocent populations, and 1) Research Goal: The goal of AI research should be to create weapons hacked to behave in undesirable ways.” not undirected intelligence, but beneficial intelligence. “We do not have long to act. Once this Pandora’s box is opened, 2) Research Funding: Investments in AI should be accompanied it will be hard to close.” by funding for research on ensuring its beneficial use, including thorny questions in computer science, economics, law, ethics, and On the other hand, the world superpowers are rapidly not only social studies, such as: developing, but also applying autonomous weapons from drones  How can we make future AI systems highly robust, so that to tanks or submarines. Some even argue that it is already too late they do what we want without malfunctioning or getting to stop the autonomous weapons hacked? Another example: the EU parliament accepted a new legislation  How can we grow our prosperity through automation while giving artificial systems some rights of live beings. This is exactly maintaining people’s resources and purpose? one of the rules of the thumb not to do to avoid the potentially  How can we update our legal systems to be more fair and negative AI progress. So, why did the EU politicians accept such efficient, to keep pace with AI, and to manage the risks a law? It is not dangerous yet, but clearly worrisome. associated with AI?  What set of values should AI be aligned with, and what legal 4. The 23 ASILOMAR PRINCIPLES and ethical status should it have? The Future of Life Institute’s 3) Science-Policy Link: There should be constructive and healthy [4] second conference on the future exchange between AI researchers and policy-makers. of artificial intelligence was organized in January 2017. The 4) Research Culture: A culture of cooperation, trust, and purpose of this paper is to present, in a rather original way as transparency should be fostered among researchers and presented at the conference, the 23 asilomar AI principles [1] developers of AI. defined at the BAI 2017 conference, accompanied with the 5) Race Avoidance: Teams developing AI systems should original discussions, the comments and analysis of the author of actively cooperate to avoid corner-cutting on safety standards. this paper. The opinion of the community is pretty a shared one: “a major 4.2 Ethics and Values change is coming, over unknown timescales but across every 6) Safety: AI systems should be safe and secure throughout their segment of society, and the people playing a part in that transition operational lifetime, and verifiably so where applicable and have a huge responsibility and opportunity to shape it for the feasible. best.” 7) Failure Transparency: If an AI system causes harm, it should The first task of the organizers was to compile a list of scores of be possible to ascertain why. opinions about what society should do to best manage AI in 8) Judicial Transparency: Any involvement by an autonomous coming decades. From this list, the organizers distilled as much as system in judicial decision-making should provide a satisfactory they could into a core set of principles that expressed some level explanation auditable by a competent human authority. of consensus. The coordinating effort was dominating the event, 9) Responsibility: Designers and builders of advanced AI systems resulting in a significantly revised version for use at the meeting. are stakeholders in the moral implications of their use, misuse, There, small breakout groups discussed subsets of the principles, and actions, with a responsibility and opportunity to shape those giving detailed refinements and commentary on them. This implications. 9 10) Value Alignment: Highly autonomous AI systems should be (e.g. the Stanford 100-year report), government (e.g. two major designed so that their goals and behaviors can be assured to align reports from the White House), industry (e.g. materials from the with human values throughout their operation. Partnership on AI), and the nonprofit sector (e.g. a major IEEE 11) Human Values: AI systems should be designed and operated report). The paper will hopefully spur discussion and awareness so as to be compatible with ideals of human dignity, rights, about these issues also in our country where it is most important freedoms, and cultural diversity. that the public, media and governance understand that the times 12) Personal Privacy: People should have the right to access, are changing fast, that new approaches and methods are needed. manage and control the data they generate, given AI systems’ Scientific comprehensions about AI, its influence on everyday power to analyze and utilize that data. life, and future for the human civilization are stacking up. 13) Liberty and Privacy: The application of AI to personal data Scientists are able to provide some guidelines in which direction must not unreasonably curtail people’s real or perceived liberty. should we humans develop AI to avoid the dangers of the 14) Shared Benefit: AI technologies should benefit and empower negative effects of the rising power of artificial intelligence. While as many people as possible. AI often frightens general public, this author finds its fast progress 15) Shared Prosperity: The economic prosperity created by AI a necessity to prevent degradation or self-destruction of human should be shared broadly, to benefit all of humanity. civilization. The potential dangers are real, not fictitious, 16) Human Control: Humans should choose how and whether to primarily to a simple fact that any major power can be easily delegate decisions to AI systems, to accomplish human-chosen misused to cause harm to humans, and second, that there are some objectives. strong indications that civilizations tend to destroy themselves. By 17) Non-subversion: The power conferred by control of highly raising awareness, we increase the chances to ripe the positive advanced AI systems should respect and improve, rather than aspects of the future mighty AI and avoid the negative ones. subvert, the social and civic processes on which the health of society depends. 6. REFERENCES 18) AI Arms Race: An arms race in lethal autonomous weapons [1] Asilomar principles. 2017, should be avoided. (https://futureoflife.org/2017/01/17/principled-ai-discussion- asilomar/). 4.3 Longer-term Issues [2] Bostrom, N. 2014. Superintelligence – Paths, Dangers, 19) Capability Caution: There being no consensus, we should Strategies. Oxford University Press, Oxford, UK. avoid strong assumptions regarding upper limits on future AI [3] Eye Scans to Detect Cancer and Alzheimer’s Disease, capabilities. https://spectrum.ieee.org/the-human- 20) Importance: Advanced AI could represent a profound change os/biomedical/diagnostics/eye-scans-to-detect-cancer-and- in the history of life on Earth, and should be planned for and alzheimers-disease managed with commensurate care and resources. [4] Future of life institute, https://futureoflife.org/ 21) Risks: Risks posed by AI systems, especially catastrophic or existential risks, must be subject to planning and mitigation [5] Gams, M. 2001. Weak intelligence: through the principle efforts commensurate with their expected impact. and paradox of multiple knowledge. Nova Science. 22) Recursive Self-Improvement: AI systems designed to [6] IJCAI conference, 2017, https://ijcai-17.or recursively self-improve or self-replicate in a manner that could [7] Kosinski, M., Wang. Y. 2017. Deep neural networks are lead to rapidly increasing quality or quantity must be subject to more accurate than humans at detecting sexual orientation strict safety and control measures. from facial images. https://osf.io/zn79k/ 23) Common Good: Superintelligence should only be developed in the service of widely shared ethical ideals, and for the benefit [8] Kurzweil, R. 2006. T he Singularity Is Near: When Humans of all humanity rather than one state or organization. Transcend Biology, Sep 26, Penguin Books. [9] Mail online, Science and technology, Vladimir Putin warns 5. CONCLUSION whoever cracks artificial intelligence will 'rule the world', http://www.dailymail.co.uk/sciencetech/article- The AI progress is already fascinating, and it speeds-up each 4844322/Putin-Leader-artificial-intelligence-rule-world.html consequent year. The rising awareness of AI-related changes in human society are appearing in scientific, academia and general [10] Yampolskiy, R.V. 2016. Artificial Superintelligence. CRC public. Dozens of major reports have emerged from academia Press. 10 Comparison of Feature Ranking Approaches for Discovery of Rare Genetic Variants Related to Multiple Sclerosis Matej Petković 1,2,B Jovan Tanevski 2 Aleš Maver 3 matej.petkovic@ijs.si Lovro Vidmar 3 Borut Peterlin 3 Sašo Džeroski 1,2 1 International Postgraduate School Jožef Stefan, Ljubljana, Slovenia 2 Department of Knowledge Technologies, Jožef Stefan Institute, Ljubljana, Slovenia 3 Clinical Institute of Medical Genetics, University Medical Centre Ljubljana, Slovenia ABSTRACT ity of the input space, so that only the features that contain In this work, we assess the quality of the ReliefF and Ge- the most information about target are kept in the dataset. By nie3 feature ranking algorithms on the task of discovering doing this, we decrease the amount of memory/time needed rare genetic variants related to multiple sclerosis using real to build a predictive model, while the performance of the world data. The data consists of a total of 183 patients model is not degraded. with multiple sclerosis and healthy controls. We evaluate the rankings and check whether two different environments Second, dimensionality reduction typically results in models for data acquisition influence the data. The results show that are easier to understand, which comes in handy when the that Genie3 algorithm produces better rankings. How- a machine learning expert works in collaboration with a ever, different environments for data acquisition have lesser domain expert. influence on the ReliefF rankings. Third, we can use FR as a guidance that reduces our search Categories and Subject Descriptors space which results in much lower costs of the subsequent experiments, for example, when we are trying to search for I.5.2 [Pattern Recognition]: Design Methodology genetic markers that indicate the presence of a disease. General Terms The last reason was the main motive for the experiments in Algorithms, Experimentation this paper. Our goal is to establish a small subset of genetic variants that can be used to learn a predictive model that Keywords accurately distinguishes between sick and healthy patients. feature ranking, genetic variants, multiple sclerosis To this end we applied two FR algorithms to the real world problem of discovery of rare genetic variants that play a role 1. INTRODUCTION in multiple sclerosis (MS) and evaluated their performance. Feature ranking (FR) is an important task in machine learn- There is a plethora of FR methods. For their overview, see ing, which can be formalized as follows. We are given a set of [8]. The result of applying a FR algorithm to a dataset, is examples x from the input domain X ⊆ X1 × X2 × · · · × XD, a score impo(xi), which tells us how much information is where D ≥ 1 is the number of descriptive attributes (fea- contained in the feature xi with regards to the target y. FR tures). We assume that the domain Xi of the i-th feature xi is then obtained by sorting the features with respect to their is either a subset of R or an arbitrary finite set, i.e., domain importance. In this work, we consider the ReliefF [6] and Xi and feature xi are either numeric or nominal. Each exam- Genie3 [4] FR algorithms. ple x is associated with a target value y(x) from the target domain Y. Given a dataset D ⊆ X × Y, the goal of FR is to The rest of the paper is organized as follows. We describe the estimate how much each of the features influences the target, considered FR algorithms in Section 2. The description of and then order the features with respect to the influences. the data and experimental design are presented in Section 3. We present the results in Section 4, and conclude in Section 5. FR is a significant part of predictive modelling. The goal of predictive modelling is to learn a model able to predict the 2. METHODS values of the target variable y, given a dataset D. The two general types of predictive modelling are regression (when In this section, the considered FR algorithms are described. Y ⊆ The section starts with the description of the ReliefF algo- R) and classification (otherwise). In our work we are concerned with classification. In classification, the values rithm followed by the description of Genie3. from Y are usually referred to as classes. 2.1 ReliefF There are three main reasons for FR with regards to predic- The motivation behind the ReliefF algorithm is the following. tive modeling. First, we may want to reduce the dimensional- Suppose the instances x1 and x2 are close to each other 11 given some distance measure, but the difference of the corre- The main motivation for Genie3 ranking is that splitting the sponding values of a feature xi is high. If x1 and x2 belong current subset E ⊆ DTRAIN, according to a test in the node to different classes, we conclude that the change of the values N where an important feature appears, should result in high of xi is one of the reasons for the change of the target value. variance reduction h(N ). Greater emphasis is put on the Hence, xi has high relevance. However, if x1 and x2 are features higher in the tree where |E| is larger. The Genie3 of the same class, then xi is not relevant, since the high importance of the feature xi is defined as difference did not cause any change of the target value. 1 X X impo (x |E(N )|h(N ), (2) GENIE3 i) = |F| The ReliefF algorithm is an iterative procedure. As can be T ∈F N ∈T (xi) seen from its pseudocode (Alg. 1), the importances of the where T (x features are stored in the list of weights w. At each of the i) is the set of nodes of the tree T where xi is part of the test and E(N ) is the set of examples that come to m iterations, we randomly select an example x ∈ DTRAIN the node N . (line 3) and find its k nearest neighbors of the same class, i.e., hits (line 4), and its k nearest neighbours from each of the opposite classes, i.e., misses (line 6). The used distance 3. EXPERIMENTAL DESIGN on the descriptive space is the sum of component distances In this section we present the data used in the experiments d that were performed to i) find the locations in human DNA i that are defined as that influence the multiple sclerosis, and ii) check how much  1[x1 6= x2 different environmental conditions in the data aggregation  i i ] : Xi nominal di(x1, x2) = |x1−x2| , (1) and processing step influences the results. i i : Xi numeric  max xi−min xi x x 3.1 Data Description At the end of each iteration, the feature importances are up- Our data consist of 183 instances corresponding to patients. dated with the weighted average of the component distances These are divided into three groups: 43 suffering from spo- between x and its neighbors. radic multiple sclerosis (SMS), 47 suffering from familial multiple sclerosis (FMS), and 93 being healthy (NoMS). The Algorithm 1 ReliefF(DTRAIN, m, k) patients are described by 202487 numeric features which 1: w ← zero list of length D describe the presence of a genetic variant in patients’ DNA, 2: for j = 1, 2, . . . , m do and the target variable which describes their diagnosis. The 3: x ← random example from D patient genomes were sequenced at the Clinical Institute of TRAIN 4: H Medical Genetics at the University Medical Centre Ljubljana. 1, . . . , Hk ← k nearest hits for x 5: for all classes c 6= xy do 6: M Based on the presence of genetic variants on the two strands c,1, . . . , Mc,k ← k nearest misses for x from c 7: for i = 1, 2, . . . , n do of DNA, as compared to a reference genome (hg19), we can 8: ⊕ ← P P (c) Pk d distinguish between three possible genotypes for every single c6=x i (Mc,l, x) /mk y 1−P (Ry ) l=1 locus: i) reference sequence on both strands, ii) presence of a 9: ← Pk d l=1 i (Hl, x) /mk genetic variant on one strand only, i.e., in heterozygous state, 10: w[i] ← w[i] + ⊕ − and iii) presence of a genetic variant on both strands, i.e., 11: return w in homozygous state. These states are respectively assigned the values 0, 1 and 2. However, the data set contains some missing values, since the success of sequencing and genotyping 2.2 Genie3 at a particular locus varies among test subjects. The Genie3 ranking is based on a forest of predictive clus- tering trees (PCTs) [1, 5] as the baseline classifiers. PCTs The feature value for the patients come from two different generalize decision trees and can be used for a variety of laboratories: the first and the second gave the results for 171 learning tasks, including clustering and different types of and 12 patients respectively. Since the different environments prediction. They are induced with the standard top-down could introduce some bias, we prepared two versions of the induction of decision trees algorithm [2], which takes a set of dataset: one containing all patients and the other, containing examples D only the patients from the first laboratory. TRAIN as input, and outputs a tree. The heuristic h that is used for selecting the tests in the tree nodes, is the reduction of variance caused by partitioning the instances in These two versions are used in the experiments where we a node of the tree. By maximizing the variance reduction, try to tell apart the three groups of patients (NoMS, SMS the homogeneity of the instances in the subbranches is max- and FMS). Following the suggestions of data providers, we imized: The algorithm is thus guided towards small trees also tried to tell apart only healthy and diseased patients. with good predictive performance. Here, we modify the target variable and merge SPS and FMS into one group (MS). The modified target can now take two To achieve better predictive performance, one can induce different values: MS and NoMS. more than one PCT and combine them into an ensemble classifier, called a forest of PCTs. The trees in the forest are 3.2 Evaluation Methodology not built on a dataset DTRAIN. Rather, different bootstrap To assess FR quality, one typically uses k-fold cross-validation replicates of DTRAIN is constructed, for each tree. The pre- (CV), where the data is divided into k parts (folds). At each diction of the forest for a given instance x is then typically of k iterations of the procedure, a ranking is constructed the class that the majority of the trees voted for. from the training set DTRAIN which is an union of k − 1 folds, 12 and then evaluated on the testing set DTEST, which is the between the sets Bj and Fj that correspond to the rankings remaining fold. At the end, the per-fold quality measures computed on a data from both laboratories (Bj ) and the are aggregated to a single ranking quality score. first laboratory (Fj ). Additionally, we also compute an ap- proximation of the expected value d JSI j of the index between This procedure is appropriate if one wants to evaluate the to random feature subsets: d J SIj = j/(2D − j). quality of a FR algorithm and is not interested in the actual FR (different FRs correspond to different training folds). This is not the case in this study. On contrary, we are 3.3 Algorithm Parametrisation interested in the quality of one particular FR that is to For the ReliefF algorithm, the default values of the pa- be reported to the domain experts, so we slightly modified rameters were used: the number of iterations was set to the standard evaluation procedure. Taking into account m = |DTRAIN|, and the number of neighbours was set to the specifics of the ReliefF and Genie3 FR algorithms, we k = 10. To compute the Genie3 ranking, a forest of 1000 adopted the following two approaches: trees is grown. The random forest subspace size was set to 25% of the features. For ReliefF, we use k-fold CV, but we do not evaluate per- fold FRs. Rather, we first average them into one single FR Since the dataset is not too big, leave-one-out CV is used by sorting the features by their average per-fold importances. for obtaining the ReliefF ranking, as well as for evaluation This average FR is then evaluated in the subsequent steps. of both average rankings. Here, the support vector machines with linear kernel were used as a classifier [3]. We use an analogous procedure for the Genie3 ranking. Note that the Genie3 importance (Eq. 2) is actually an average 4. RESULTS AND DISCUSSION of importances for different trees in the forest. Moreover, To asses, which of the FRs found more promising genetic each tree is built on different bootstrap replicate of the data markers in human DNA that influence the MS, we compute which does not contain all known examples. This is why the feature addition curves (Sec. 3.2). Fig. 1 shows the results we simply run the algorithm on the whole dataset D. The for the first laboratory and binary target, but the graphs for obtained ranking is then evaluated in the subsequent steps. the other three versions of the data are similar. The remainder of the evaluation procedure is the same for More specifically, in all four cases the ranking algorithms both FR methods and is a variant of the one proposed by successfully discover important features, since the curves are Slavkov [7]. Again, we use k-fold CV. At each iteration, we ascending in the first part when relevant features are added first build a predictive model on a training fold DTRAIN, con- to feature subsets. Later on, the irrelevant features prevail sidering only the j topmost features of the average ranking, and the performance slowly decreases (see Fig. 1). Next, at for each value of 1 ≤ j ≤ D, such that j = 1 + `(` + 1)/2 for the beginning, Genie3’s curve is always clearly above the somè ∈ N or j = D. Each of these models is then tested ReliefF’s. Finally, the maximal accuracy of Genie3’s ranking on the testing fold DTEST. is always higher than ReliefF’s, and is also achieved sooner. The result of cross-validation are confusion matrices Mj . The As a consequence, the α scores of the Genie3 FRs are higher (c, d)-th entry of the matrix Mj tells how many patients from than those of ReliefF, as shown in Tab. 1. This table also the class c were assigned the class d by the classifier that was shows that the data coming from both laboratories and built from the topmost j features in the ranking. having binary target, result both in the best FR among all Genie3 FRs, and in the worst FR among all ReliefF FRs. Let αj denote the accuracy, computed from the matrix Mj . The points (j, αj ) form a feature addition curve. The mo- tivation behind this approach is that for higher αj ’s, more Table 1: The α scores of the Genie3 and ReliefF relevant features are positioned at the beginning of the FR. rankings, for all versions of the data. Moreover, from the shape of the curve, we can deduce some Laboratory Target values Genie3 ReliefF qualitative characteristics of the FR. E.g., if the curve does Both {MS, NoMS} 0.804 0.633 not ascend in some part, that means only redundant or ir- Both {FMS, SMS, NoMS} 0.697 0.677 relevant features are placed in the corresponding part of the First {MS, NoMS} 0.753 0.605 FR. First {FMS, SMS, NoMS} 0.745 0.711 If we want to express the quality of a FR as a single number, We inspect the influence of different sources of data by com- we can compute the weighted average α of the accuracies αj : puting the JSI (Eq. 3) between the sets of the topmost α = (P w w j j αj )/w, where w = Pj j and the weights wj features of the FRs that base on data from both laboratories decrease with j, since the beginning of a FR is considered the and from the first laboratory only. Fig. 2 shows that different most important. In our experiments, we choose wj = 1/j, as sources notably influence the FRs. The fluctuations at the suggested by Slavkov [7]. A good FR has a high α score. very beginning are expected, since every difference greatly influences the JSI values, when feature subsets are small. To asses the influence of different sources of the data, i.e., two laboratories, we use the Jaccard similarity index. For a After that, the curve of the ReliefF ranking stabilizes at fixed size j of the set of topmost features, we compute the approximately 0.8 which means that these rankings identify Jaccard similarity index the same features as important. This does not hold in the JSI j = |Bj ∩ Fj |/|Bj ∪ Fj | (3) case of Genie3 rankings. The corresponding sets of 10012 13 1.0 1.0 0.9 0.9 0.8 0.8 accuracy accuracy 0.7 0.7 0.6 Genie3 (407: 1.0) 0.6 Genie3 (407: 1.0) ReliefF (3829: 0.971) ReliefF (3829: 0.971) 0.5 0.5 0 50000 100000 150000 200000 0 1000 2000 3000 4000 number of features number of features Figure 1: Feature addition curves for the rankings produced by the Genie3 and ReliefF algorithms, using data from the first laboratory only and considering a binary target (left: complete feature addition curves, right: feature addition curves for the first 4000 features). The numbers in the brackets correspond to the maximum accuracy and the number of features where it is first achieved. 1.0 and analyzed a small subset of relevant features. They com- pared the top ranked features to results reported in the 0.8 literature. In the small subset they found matches to genes that have been reported to be associated with MS. Given the positive matching, additional experimental validation of 0.6 top ranked features can be performed in order to determine JSI the existence of previously unconsidered causal relations. 0.4 Genie3 We plan to run the algorithms on a new version of the 0.2 data, processed using a new pipeline that takes into account ReliefF the different environmental conditions. Since there is a expected similarity 0.0 taxonomic relation between the classes, we will also consider hierarchical classification as the baseline for FR. 0 50000 100000 150000 200000 number of features 6. REFERENCES [1] H. Blockeel. Top-Down Induction of First Order Logical Figure 2: Jaccard similarity of the topmost features Decision Trees. PhD thesis, Katholieke Universiteit of the rankings using the data from the first and Leuven, 1998. both laboratories. The expected similarity corre- sponds to random ranking and serves as a baseline. [2] L. Breiman, J. Friedman, R. Olshen, and C. J. Stone. Classification and Regression Trees. Chapman & Hall/CRC, 1984. features still have [3] I. Guyon, J. Weston, S. Barnhill, and V. Vapnik. Gene d JSI < 0.2, which means that different features are recognized as important. Therefore, the Genie3 selection for cancer classification using support vector ranking is more sensitive to changes in the data, since the machines. Machine Learning, 46(1):389–422, 2002. 12 additional patients from the second laboratory notably [4] V. A. Huynh-Thu, L. Irrthum, Wehenkel, and P. Geurts. changed the rankings. This finding also confirms the data Inferring regulatory networks from expression data using providers’ concerns about the influence of different environ- tree-based methods. PLoS One, 5(9):1–10, 2010. mental conditions on the data aggregation and processing. [5] D. Kocev, C. Vens, J. Struyf, and S. Džeroski. Tree ensembles for predicting structured outputs. Pattern 5. CONCLUSIONS Recognition, 46(3):817–833, 2013. We used the Genie3 and ReliefF algorithm to identify rare [6] I. Kononenko and M. Robnik-Šikonja. Theoretical and genetic variants related to multiple sclerosis. The feature Empirical Analysis of ReliefF and RReliefF. Machine addition curves reveal that the rankings produced by the Learning Journal, 55:23–69, 2003. Genie3 algorithm are better than those of ReliefF, but they [7] I. Slavkov. An Evaluation Method for Feature Rankings. are also more sensitive to changes in the data, as shown by PhD thesis, International Postgradudate School Jožef low values of the JSI score. Stefan, Ljubljana, 2012. [8] U. Stańczyk and L. C. Jain, editors. Feature Selection However, since the Genie3 algorithm consistently outper- for Data and Pattern Recognition. Studies in formed ReliefF in terms of α scores, only the Genie3 rankings Computational Intelligence. Springer, 2015. were reported to the domain experts. They further focused 14 Modeling of Dynamical Systems: A Survey of Tools and a Case StudY Gjorgi Peev1,2 Nikola Simidjievski1 Sašo Džeroski1,2 1Jozef Stefan Institute 2Jozef Stefan International Postgraduate School Jamova cesta 39, Ljubljana, Slovenia (Emails: Name.Surname@ijs.si) presents a challenge when it comes to visualizing the structure of ABSTRACT the modeled system. While scientists can typically comprehend and Process-based modeling refers to an approach for automated relate to models formalized as equations, the (uncommon) high- construction of models of dynamical systems from knowledge and level PBM formalism is not always familiar to them. Currently, this measurements. The underlying formalism allows for both makes ProBMoT usable for a narrow scope of domain experts. On explanatory representation of a dynamical systems in terms of the other hand, several state-of-the art grey-box modeling software principle system components, as well as their transformation into such as: Prometheus [6], Eureqa [7], MATLAB [8], STELLA [9] equations adequate for simulation. The process-based modeling and COPASI [10] have been used extensively for different approach, while successful in addressing a variety of modeling modeling tasks in a variety of domains. tasks, still struggles with meeting some user-interface criteria In order to widen ProBMoT’s user base, in this paper we aim at acceptable for a wider scope of users. In this paper, we review identifying the main features and limitations of each of the several state-of-the-art approaches and formalisms for (automated) aforementioned modeling software and compare them to modeling of dynamical systems, and compare them to the most ProBMoT. In particular, we attempt at modeling a two-cascaded recent implementation of the process-based modeling approach – water tanks system, a well-known system identification ProBMoT (Process-based modeling tool). benchmark, with each of the six modeling tools and compare them Keywords in terms of their input and output according to several criteria. automated modeling, process-based modeling, dynamical systems, However, quantifying and describing a modeling software and its formalism, software formalism, is not a trivial task. To this end, we propose five criteria according to which we survey the different modeling approaches: (C1) Generality: the applicability of a software to a general 1. INTRODUCTION problem (from all fields). In contrast, there are software applicable Models of dynamical systems yield a mathematical representation to problems from specific fields (molecular biology, finances, of the nature laws that govern the behavior of the system at hand. ecology, electronics, etc.). Such models are employed to recreate or simulate the behavior of (C2) Parameter estimation: capability of fitting the model’s dynamical systems under diverse conditions. parameter values to data. The two principal elements of every approach to modeling (C3) Automated modeling: ability to learn models with automated dynamical systems are (1) structure identification and (2) parameter computational scientific discovery methods. Note that, here we can estimation. The former tackles the task of establishing a structure distinguish also between fully automatic and semi-automatic of a model in terms of equations, while the latter deals with approaches. The former does not rely on prior knowledge about a approximation of the constant parameters and initial values of the domain and typically results in one model structure built from variables in the model for a given structure. Typically, two scratch. The latter refers to the ability to discover a set of approaches are being used for modeling dynamical systems: explanatory models blending expert domain knowledge with knowledge-driven (white-box) and data-driven (black-box) computational discovery algorithms. modeling. The former relates to a domain expert deriving a proper (C4) Graphical representation: ability to graphically represent the structure of a model by employing extensive knowledge about the output models of the software. system at hand. In turn, the model’s parameters are estimated either (C5) Comprehensibility: whether the output of the software is by using measured data, or manually based on the expert’s comprehensible on first hand to the domain-expert user, without the experience. The latter methodology refers to a trial-error principle: need of background knowledge. it uses measured data to search for a structure/parameters The rest of this paper is organized as follows. In the next section combination that best fits the observed behavior. we outline the six state-of-the-art tools for modeling dynamical Process-based modeling (PBM) [1,2,3], refers to a grey-box systems. Section 3 elaborates the design of the modeling approach, since it joins the knowledge- and data-driven modeling experiment and presents the results, which in turn are discussed in approaches and allows for automated modeling of dynamical Section 4. Finally, Section 5 concludes the paper. systems. In particular, process-based modeling employs both domain-specific knowledge and data for simultaneously 2. BACKGROUND constructing the structure of the model and estimating its parameters. The resulting process-based model offers both high- In this section, we focus on six grey-box modeling software level explanatory representation of a dynamical systems in terms of packages and their characteristics: ProBMoT, Prometheus, Eureqa, its principle system components, as well as their transformation MATLAB, STELLA and COPASI. into a low-level formalism in terms of equations adequate for ProBMoT [4,5] (Process-Based Modeling Tool) is the latest simulation of the system’s behavior. implementation of the process-based modeling paradigm. It is a The latest implementation of the process-based modeling paradigm software for construction, parameter estimation and simulation of – ProBMoT [4,5] uses text-based, non-visual formalism, which process-based models. 15 The output of this software is a process-based model, represented to STELLA is a user-specified quantitative model and measured with entities and processes. Entities relate to the actors of the data. Similarly, STELLA also does not support automated observed system, defined with constants and variables. The modeling, therefore the output model structure is never learned. processes, on the other hand, represent the interactions between the However, one can still simulate the complete model by invoking entities, referring to one or more ordinary/algebraic equation. the simulator that can run the input user-defined model. Collating the equations from all the processes in the model, a COPASI [10] is a software for simulation and analysis the system of ODEs can be attained. dynamics of biochemical networks. It supports models in the To this end, ProBMoT takes as input a library of domain SBML standard [12]. The models are defined with chemical knowledge, a task specification and data. reactions between molecular species. They can also include The library is formalized by establishing templates of generic compartments, events, and other global variables that can help entities that appear in the generic processes. The templates can be specify the dynamics of the system. Here, we can also draw an organized into a hierarchical structure. The task specification limits analogy between COPASI’s and ProBMoT’s formalisms. The the search space of candidate model structures by supplying species in COPASI correspond to ProBMoT’s entities, while constraints, specified as incomplete conceptual models as modeling reactions are analogous to processes containing equations which presumptions. The library of domain-specific knowledge together describe the behavior of the system. However, in contrast to with the task specification determine the space of models. The ProBMoT, COPASI does not perform automatic structure induction algorithm then searches through this space of candidate identification. The input to COPASI is user-defined model and model structures, finding plausible model solutions and estimating measured data. The result is a complete model, with parameters the constant parameters of each candidate model structure to the tuned to best fit the data. input data. Prometheus [6] is a software that supports interaction between the 3. CASE STUDY user and computational discovery algorithms. The formalism used In order to better illustrate and evaluate the formalisms of the in Prometheus specifies process models and background software described in the previous section, here we tackle the task knowledge in terms of variables and processes that relate them. of modeling a two-cascaded water tanks system [13]. The system Each process express casual relations between its input and output is consisted of two cascaded water tanks with free outlets, fed by a variables through one or more differential equations. pump. The governing equations for this system are depicted below The input for this software is a user-defined model, library of (Eq. 1), where the states of the water levels of the two tanks are background knowledge consisted of generic processes, measured denoted with h data and constraints specifying what can be revised. The output is 1 and h2, the latter ( h2) being the output. The voltage applied to the pump is u(t), while A a revised model-structure that best fits the measured data. 1, A2, a1 and a2 denote the areas of the tanks and their effluent areas, while the applied voltage-to- Prometheus is a predecessor to ProBMoT, and consequently their flow conversion constant is denoted with k. The task is to model the formalisms are comparable. Prometheus uses process models, response of the lower tank. which are analogous to ProBMoT’s process-based models. Their definitions for processes as model’s components are similar. The difference is that the variables in Prometheus are not encoded in an entity, but they are represented as a component. Eureqa [7] uses symbolic regression [11] with genetic programming in order to infer equation-based structure of the Equation 1. Two-cascaded water tanks system system and its parameters solely from data, by minimizing the error using the implicit derivatives method. The state of the modeled ProBMoT - In order to model the water tanks system in ProBMoT, system is declared with a target variable, its descriptors and their we first need to create a library of domain knowledge (Figure 1A), form in order to define the search space. Note that, the modelers i.e. we need to formulate template entities and processes which in have little-to-no control over the space of plausible structures. This turn will be instantiated to specific entities and processes. The main means that, it is still a domain expert’s task to infer the similarities actors in the system are two water tanks (with same properties) and between the resulting model and the real system structure. a pump. In terms of template entities, this translates to one template MATLAB [8] supplies functions for performing system entity Thank and a template entity Pump. The tanks are identification and parameter tuning of a user pre-defined model. Its characterized with a variable h, representing the water level height, formalism allows specifying quantitative models with and a constant outflow_c which denotes the ration between the instantaneous and differential equations. Note that, MATLAB does tanks areas (a/A). The dynamics that govern the system’s behavior not support automated modeling, i.e., learning multiple structures. in terms of equations are encoded in template processes inflow, The models are defined as model objects, i.e. specialized data containers that encapsulate data and other model attributes. The ValveTransmission, outflow. These correspond to the water inflow dynamics of the system at hand are described with ODEs imported in the first tanks, the water flows between the two tanks, and the in a C MEX-file. With associating the model object to the C MEX- water outflow from the second tank, respectively. In terms of file, and employing functions for simulation and parameter defining modeling constraints, we can outline the number of estimation, on the output MATLAB obtains a completely defined entities involved in the system and encode the plausible process model structure with all parameters tuned in accordance to data. alternatives. In turn, such a library together with the modeling STELLA’s [9] formalism relates to stocks, flows, convertors and constraints, can be induced to a specific model structure (Figure connectors. Stocks represent variables, flows denote their changes 1B-top) of the particular system, parameters of which are fitted over time (derivatives), converters encode the constant parameters, using the measured data. Such a model can be then transformed while connectors are used to attain a link between all of them. into to a system of ODEs (Figure 1B-bottom) and simulated. While the models are built using this formalism, the software produces finite difference equations that describe it. Note that, this formalism is comparable to the PBM. We can associate stocks with entities, and flows with processes. Similar to MATLAB, the input 16 MATLAB - We describe the dynamics of the two-tank system with writing the two differential equations that govern its behavior into a C MEX-file. First, we initialize all the parameters that we are going to use for modeling. Second, we write a function ( compute_dx), which computes the state equations, i.e. the change 𝑑ℎ1 𝑑ℎ2 in the water height level of the two tanks over time ( and , 𝑑𝑡 𝑑𝑡 represented as dx[0] and dx[1] respectively). Next, we write a function ( compute_y), which computes the output equation, in our case the response of the lower tank, i.e. y[0] = x[1]. In turn, with the Identify non-linear grey box model function in MATLAB, we associate a model object with the C MEX-file, resulting in a grey- box model of the system at hand (Figure 4). We choose which of A) B) the parameters described we want to estimate, and with the function Figure 1. A) Library of background knowledge for a water Non-linear grey-box estimate they are fitted to the data provided at tank system B) Process-based model of a particular water- input. The obtained model can then be simulated. tank system (top); The same model transformed to ODEs (bottom) Prometheus - In a similar fashion to ProBMoT, we define processes with equations in Prometheus as well. We have process inflow, valvetransmition and outflow. In this formalism, entity components are not present, but the variables represent a component themselves. Consequently, we have three variable components: h1, h2 (as observable), and v (as exogenous). We create a library of background knowledge, containing generic Figure 4. Structure of a C MEX-file with and output equation processes, where the equations instead of numbers have and estimated parameters parameters, and with the measured data, we refine the model structure. We obtain the final defined model structure with STELLA - We model the particular system at hand with two stocks estimated parameters. The models obtained in Prometheus (Figure as the main actors of the system, representing the two water tanks. 2) are highly comprehensible, since the software has visual The change in their water height is indicated with three flows. Flow representation for its models. #1 represents the amount of water transferred from the upper to the lower tank: the outflow of the upper is an inflow for the lower tank. Flow #2 is the outflow from the lower tank and Flow #3 is inflow from the pump into the upper tank. The other components and parameters are represented with convertors. The model obtained in STELLA is graphically highly comprehensible. The drawback is that this software does not support automated parameter estimation. Figure 2. Prometheus model of a particular water tank system 𝑑ℎ2 Eureqa - For the two-tank system, we define as the target Figure 5. STELLA model of a particular water tank system 𝑑𝑡 variable. The form of the equation that describes our target should COPASI - We compose our model with three compartments, 𝑑ℎ1 use h1, h2, and v as descriptive variables. During the search, 𝑑𝑡 species and reactions (Figure 6). The lower, the upper tank and the we can visualize (Figure 3) the found equations, ranked on a environment (in which the two tanks are in) represent different complexity-error graph. If the error is not decreasing significantly, compartments themselves. Compartment Tank1 has initial and the complexity is increasing, we can stop the search at any time. expression of the area of the upper tank ( A1). Similarly, Tank2 has In the end, it results with a set of equations, from which we choose initial expression of the area of the lower tank ( A2). In similar the most suitable one, obtaining a complete model for our water fashion to ProBMoT, where each entity has variables or constants, tanks system. here each compartment contains species: Tank1 includes height h1, Tank2 holds height h2 and Environment contains the voltage u applied to the pump. Reactions in COPASI are analogous to ProBMoT’s processes, and knowing the equations from (Eq.1), we define reaction flow (between h1 and h2) reaction inflow (between u and h1) and reaction outflow (from h2) . As an output, we obtain a model of differential equations with computed parameter values. COPASI is widely used in the field of biochemical networks and their dynamics. However, using COPASI outside of that fields, as Figure 3. Eureqa model of a particular water tank system the case in this paper, is not a trivial task. 17 two-water-tanks system with all the different tools and compare them in terms of their input and output according to five criteria. The general conclusion of this paper is that ProBMoT, the latest software implementation of the PBM paradigm, while successful in tackling various modeling tasks, is usable for a very narrow scope . of users mainly because of its uncommon high-level modeling Figure 6. COPASI model of a particular water tank system language and the lack of graphical representation of the resulting models. We conjecture that establishing a Graphical User Interface 4. DISCUSSION (GUI) for it, will address its usability issues. However, in order to create a GUI for ProBMoT, we first need to Having the identified the key characteristics of each of the six address the non-trivial and abstract problems of visually modeling software, here we compare them based on the criteria representing the hierarchical nature of the process-based models. defined in Section 1. Table 1 presents the results of the study. One answer could be presenting the components of the process- based models as building blocks, similar to the model components Table 1. Comparing the different modeling tools according the employed in STELLA or Prometheus. It would enable the user to five different cirteria make libraries and define tasks graphically and interactively. C1 C2 C3 C4 C5 Another feature that could be of good use for the GUI is tightly ProBMoT      Prometheus      connected with the runtime process, the results and data Eureqa      visualization. Similarly to Eureqa, during the search, all found MATLAB      feasible models can be listed and ranked. STELLA      To conclude, with developing a self-explanatory visual COPASI      representation of the process-based modeling formalism, Based on the first criterion (C1 - generality) all software except comprehensible for domain-expert scientists, the PBM paradigm COPASI are general-purpose, meaning that systems from different would become more approachable. With a universal visual fields can be easily modeled and simulated with them. COPASI is representation, scientists from different fields would be able to specific-purpose software, built around the logic of the biochemical transfer knowledge between them. networks and their dynamics. This means building every type of Acknowledgements. We would like to acknowledge the support of prof. model in COPASI can be very challenging and ambitious. dr. Ljupco Todorovski. According to the second criterion (C2 - parameter estimation), all tools except STELLA have integrated parameter estimation 6. REFERENCES methods. In STELLA, the parameters are tuned manually. [1] Džeroski, S., Todorovski, L. (2003). Learning population In terms of automated modeling (C3), ProBMoT, Prometheus and dynamics models from data and domain knowledge. Ecological Eureqa are capable of automated modeling. Eureqa is able to infer Modelling, 170(2-3):129-140, Elsevier. equation-based model from scratch using genetic programing. On [2] Langley, P., Saanchez, J., Todorovski, L., Džeroski, S. (2002). the other hand, ProBMoT and Prometheus are able to find set of Inducing process models from continuous data. Proc. 19th models with similar structure components and distinguish among International Conference on Machine Learning 2002, pp. 347- 354 them. Both of them relay on domain-specific modeling knowledge [3] Todorovski, L., Džeroski, S. (2006). Integrating knowledge-driven used in the process of induction of model structures. and data-driven approaches to modeling. Ecological Modelling, In terms of graphical representation (C4), only STELLA and 194(1-3):3-13, Elsevier. [4] Čerepnalkoski, D. (2013). Process Prometheus have the ability to graphically visualize the models and -based models of dynamical systems: Representation and induction, PhD Thesis, Jožef Stefan their components, in the form of building blocks. International Postgraduate School, Ljubljana, Slovenia. Finally, in terms of comprehensibility (C5), the high-level [5] Tanevski, J., Simidjievski, N., Todorovski, L., Dzerovski, S. modeling formalisms used by ProBMoT and Prometheus results in (2017). Process-based Modeling and Design of Dynamical Systems. not widely interpretable models. Still, both have the ability to Proc. Machine Learning and Knowledge Discovery in Databases: transform the high-level formalism in equations, with additionally European Conference, ECML PKDD 2017. Prometheus having the ability to visualize the modeling [6] Sanchez, J.N., Langley, P. (2003). An interactive environment for components. While MATLAB is widely used for modeling tasks, scientific model construction, Proc. 2nd international conference on its formalism still requires a low-level programing knowledge for Knowledge capture, K-CAP’03, pp. 138-145, ACM. one to be able to encode and decode the models. Regarding Eureqa, [7] Schmidt, M., Lipson, H. (2009). Distilling Free-Form Natural while the output models are contained of (differential) equations, Laws from Experimental Data. Science 324(5923):81-85, AAAS. the model structure doesn’t necessarily correspond to the real [8] MathWorks, Inc. (2004). MATLAB: the language of technical system’s structure. The obtained models in COPASI are computing: computation, visualization, programming. USA. comprehensible for experts in the field of biochemical networks. [9] Richmond, B. (1985). STELLA: Software for Bringing System Finally, the models obtained from STELLA besides being offering Dynamics to the Other 98%. Proc. 3rd International Conference of the graphical visualization the models can also be translated into System Dynamics Society, pp. 706–718 equations. Either way, it is highly interpretable. [10] Hoops, S., Sahle, S., et. al. (2006). COPASI: a COmplex PAthway SImulator. Bioinformatics 22(24):3067-3074 [11] Koza, J.R. (1992). Genetic Programming: On the programming 5. CONCLUSION of computers by means of natural selection. MIT Press, USA In this paper, we give an overview of a formalism for automated [12] Hucka, M., Finney, A., et. al. (2003). The systems biology modeling of dynamical systems, named process-based modeling. It markup language (SBML): A medium for representation and is not always easily interpretable, making it usable only for a exchange of biochemical network models. Bioinformatics 19 narrow scope of domain experts. In order to improve that, we (4):524–531, Oxford University Press. review several state-of-the art grey-box modeling tools, identifying [13] Wigren, T., Schoukens, J. (2013). Three free data sets for development and benchmark in nonlinear system identification. Proc. their main features and limitations. As a case study, we model a European Control Conference 2013, pp. 2933–2938, IEEE. 18 Vpliv različnega prenosnega kanala pri referenčnih in testnih posnetkih na forenzično verifikacijo govorcev Tomaž Šef Robert Blatnik Institut “Jožef Stefan” Institut “Jožef Stefan” Jamova cesta 39 Jamova cesta 39 1000 Ljubljana 1000 Ljubljana +386 1 477 34 19 +386 1 477 32 69 tomaz.sef@ijs.si robert.blatnik@ijs.si POVZETEK Forenzični pogoji so doseženi, ko se dejavniki variabilnosti, ki predstavljajo t.i. »stvarne pogoje«, pojavljajo brez kakršnegakoli V članku obravnavamo problematiko razpoznavanja oz. principa, pravila ali norme. Lahko so konstantni preko celotnega verifikacije govorcev v forenzične namene in vpliv različnih klica ali pa se hipoma pojavijo ali izginejo; na celoten proces načinov zajemanja govornega signala na rezultate izvedenih analiz. vplivajo povsem nepredvidljivo. Izvedli smo poizkuse s pomočjo komercialnega sistema za samodejno razpoznavanje govorcev (SRG) in preučevali razlike v 2. METODE IDENTIFIKACIJE OZ. njegovi uspešnosti glede na različne kombinacije prenosnih VERIFIKACIJE GOVORCEV kanalov pri zajemanju učnih oz. referenčnih in testnih posnetkov. Različne metode identifikacije govorcev so lahko bolj ali manj Podani so rezultati eksperimentov za slovenske govorce, ki smo jih subjektivne oz. objektivne. Tudi pri objektivnih metodah imamo simultano snemali preko petih različnih prenosnih kanalov in ob opraviti z določenim vplivom človeka; npr. računalnik je treh različnih načinih govorjenja: branje, spontani govor in dialog. sprogramiran, rezultati pa so interpretirani s strani eksperta. Najbolj subjektivna metoda identifikacije govorcev v forenzične namene je Ključne besede slušno-zaznavna metoda oz. slušna analiza. Nekoliko bolj Forenzične analize, razpoznavanje oz. verifikacija govorcev, objektivna je slušno-instrumentalna metoda. Med najbolj prenosni kanali za zajemanje govora, govorna zbirka. objektivne štejemo polavtomatske in avtomatske metode identifikacije govorcev. 1. UVOD Slušno-zaznavna metoda (angl. »aural-perceptual approach«) oz. Pri razpoznavanju oz. verifikaciji govorcev v forenzične namene slušna analiza (angl. »auditory analysis«) v osnovi temelji na imamo opravka s spornimi posnetki izgovarjav, ki predstavljajo pozornem poslušanju posnetkov s strani izkušenega fonetika, pri dokazno gradivo in so posneti v »stvarnih pogojih« med samim čemer se zaznane razlike v govoru uporabijo za ocenjevanje stopnje izvajanjem kaznivih dejanj. V večini primerov govorni posnetki podobnosti med glasovi. Slušna analiza ima svoje omejitve in se pri predstavljajo telefonske pogovore, pridobljene predvsem na dva običajni fonetični analizi uporablja predvsem za izluščenje načina: (i) anonimen klic, kadar je pričakovan ali kako drugače zanimivih lastnosti in parametrov, ki jih nato podrobneje dostopen, (ii) prisluškovanje telefonskim pogovorom s strani analiziramo s slušno-instrumentalno metodo [2, 3]. policije. Pojem »stvarni pogoji« uporabljamo kot nasprotje Slušno »laboratorijskim pogojem«, ko ne moremo nadzirati, pričakovati -instrumentalna metoda (angl. »auditory instrumental approach) vključuje meritve različnih parametrov, kot so npr. ali predvidevati pogojev v katerih se bodo pridobili posamezni govorni posnetki. Celo več; obtoženec ponavadi ne želi korektno osnovna frekvenca (F0), hitrost govora, potek osnovnega tona, sodelovati in skuša ovirati ali preprečiti pridobitev kakršnihkoli razne spektralne karakteristike govornega signala itd. Parametri se nato medsebojno primerjajo po srednjih ali povprečnih vrednostih zanj obremenilnih informacij. in variancah. Pri računalniški akustični analizi (angl. Zaradi »stvarnih pogojev« pridobivanja posnetkov je govorni »computerised acoustic analysis) dobimo numerične vrednosti signal bolj spremenljiv oz. variabilen. Vire variabilnosti govornega različnih govornih parametrov s pomočjo posebne programske signala lahko razvrstimo v naslednje kategorije [1]: opreme. Pri tem je vloga eksperta še vedno zelo pomembna, saj se je potrebno odločiti, kateri govorni vzorci so dovolj dobre kvalitete (i) svojske variabilnosti govornih signalov istega govorca: vrsta govora, staranje, časovni presledek med dvema posnetkoma, za analizo. Poleg tega je potrebno izbrati oz. določiti primerljive narečje, žargon, socialni status, čustveno stanje, uporaba dele govornih vzorcev, ki bodo analizirani, in ovrednotiti dobljene rezultate. Parametri pri akustično forenzični analizi večinoma omamnih sredstev itd. izvirajo iz lingvistično-fonetičnih raziskav in so neposredno (ii) izsiljene oz. umetne variabilnosti govornih signalov istega povezani s slišnimi fonetičnimi značilnostmi [4]. govorca: »Lombardov« učinek, stres zaradi zunanjega vpliva, »cocktail-party« učinek itd. Polavtomatsko (angl. »forensic semiautomatic speaker recognition«) in avtomatsko (angl. »forensic automatic speaker (iii) zunanja variabilnost odvisna od kanala: tip telefona ali recognition«) razpoznavanje govorcev v forenzične namene je mikrofona, fiksna/mobilna telefonija, komunikacijski kanal, pasovna širina, dinamični obseg oz. razpon, električni in uveljavljen termin za metode (pol)avtomatskega razpoznavanja akustični šum, odmev, popačenje itd. govorcev, ki so prilagojene za uporabo v forenzične namene. Pri polavtomatskih metodah prihaja med preiskavo do interakcije eksperta in računalnika. Pri avtomatskem razpoznavanju govorcev 19 pa se medsebojno primerjajo statistični modeli akustičnih jezikoslovne vsebine in trajanja posnetkov lahko vpliva na parametrov glasov znanih govorcev (iz govorne baze) s statističnim zanesljivost dobljenih rezultatov [5]. modelom akustičnih parametrov nepoznane osebe, ki jo želimo Zadnje čase temeljijo sistemi za (pol)avtomatsko razpoznavanje identificirati (slika 1). Na podlagi te primerjave izračunamo govorcev v forenzične namene na oceni kvocienta verjetnosti (angl. kvantitativno oceno podobnosti med (od govorca odvisnimi) »likelihood ratio«; LR) [6]. Razmerje verjetnosti (LR) je podano parametri glasu nepoznane osebe na posnetku in parametri obdolženca s čimer ocenimo preprič kot razmerje gostote verjetnosti porazdelitev razlik znotraj vzorcev ljivost dokaza. Pri osumljenca in porazdelitev razlik glede na vzorce populacije v točki avtomatskem razpoznavanju govorcev (slika 2) v forenzične namene je prepričljivost dokaza odvisna od relativne verjetnosti, da E (dobimo jo s primerjavo preiskovanega posnetka in statističnega opazimo neke značilnosti nepoznanega glasu v statističnem modelu modela osumljenca) [5]. akustičnih parametrov obdolženca in v statističnih modelih glasov Metode razpoznavanja govorcev, ki temeljijo na tehnikah potencialne populacije. statističnega modeliranja, kot npr. Gaussov mešani model (angl. Gaussian Mixture Modell, GMM), imajo to dobro lastnost, da Podatkovno vodena Bayesova metoda za avtomatsko neposredno vrnejo verjetnost, ali posamezna izgovorjava lahko razpoznavanja govorcev zahteva poleg preiskovanega posnetka pripada statističnemu modelu govorca. Namesto GMM lahko za (oz. sledi) uporabo še treh baz izgovarjav (slika 1): referenčno govorno bazo osumljenca (R), ki služi izdelavi statističnega modela razpoznavanje govorcev uporabimo tudi prikrite Markovove modele ali nevronske mreže [5]. So pa te metode manj uporabne v njegovega glasu (pogoji snemanja morajo biti čim bolj podobni forenzičnih postopkih, ker nam določene verjetnosti v praksi pogojem pri snemanju govorne baze populacije P), kontrolno praviloma niso poznane in posledično ne moremo izračunati govorno bazo osumljenca (C), ki služi ocenjevanju notranje razmerja verjetnosti (LR), ki je najpogosteje uporabljan oz. edini variabilnosti glasu osumljenca (pogoji snemanja morajo biti čim sprejemljiv način podajanja rezultatov na sodiščih [6]. bolj podobni pogojem snemanja preiskovanega posnetka) in govorno bazo potencialne populacije (P), ki vsebuje takšne Avtomatski sistemi za razpoznavanje govorcev se ne smejo posnetke glasov, da nobeden naključno izbran posnetek iz te baze uporabljati samostojno pač pa le kot dopolnitev drugih metod; sicer ni izgovorjen s strani iste osebe, kot je preiskovani posnetek (oz. obstaja možnost napačne identifikacije [7]. Rezultate različnih sled). Kakršno koli neujemanje govornih baz zaradi okoliščin pri forenzičnih metod v praksi preučujemo povezano s čimer dobimo prenosu govornega signala, vrste snemalne naprave, šuma, kombinirano oceno zanesljivosti dokaznega gradiva. Slika 1. Shematski prikaz izračuna razmerja verjetnosti (LR) pri razpoznavanju govorcev. avtomatsko razpoznavanje govorcev izboljšave izluščenje modeliranje A/D govornega govornih govornih računanje signala razdalj značilk značilk Slika 2. Postopek avtomatskega razpoznavanja govorcev. 20 smo priložili 87 posnetkov pogovorov slovenskih moških. V teh 3. GOVORNA ZBIRKA posnetkih govori približno 10 različnih oseb, ki so posnete prek Glede na omejeno količino urejenih in tehnično primernih govornih prisluhov v mobilnem telefonskem omrežju. posnetkov v slovenskem jeziku smo se odločili za izvedbo Učenje modelov govorcev je potekalo s posnetki, posnetimi preko snemanja lastne govorne zbirke slovenskih govorcev. vseh petih kanalov. Učenje s posnetki prek mikrofona, smo izvajali Govorno zbirko smo posneli v laboratoriju. Posneli smo 25 moških s posnetki prek obeh mikrofonov pri branju in spontanem govoru, slovensko govorečih oseb. Izbrali smo govorce različnih starosti, torej skupno s štiri mi posnetki za vsakega govorca. Sistem smo učili tudi s posnetki vse delovno aktivne, v starostni skupini od približno 25 do 65 let. prek telefonije GSM, PSTN in VoIP za vseh izbranih 18 govorcev iz lastne govorne zbirke (branje in spontani Večji del posnetkov so govorci hkrati govorili v dva namizna govor). mikrofona (v oddaljenosti 15 do 30 cm od ust govorca), v prostoročni mikrofon VoIP telefona ter v GSM telefon in slušalko Testiranje smo izvajali s posnetki iz lastne govorne zbirke prek klasičnega analognega PSTN telefona. Na ta način smo isti govor vseh petih kanalov, ki so bili posneti pri pogovoru. posneli preko več sočasnih prenosnih kanalov, kar nam omogoča Za učenje smo uporabljali eno vrsto kanala, testiranje pa smo analize vplivov kanala na istem izvornem govornem signalu. izvedli na podatkih posnetih prek istega in vseh preostalih kanalov. Tako smo lahko opazovali obnašanje sistema pri istem modelu Govorci so pod nazorom operaterja snemanja govorili na tri načine: ozadja, vendar pri različnih kanalih za učenje in testiranje. Na spontani govor, pogovor in branje. slikah 2, 3, 4 in 5 so prikazani rezultati štirih sklopov meritev Vsak način govora smo posneli v dolžini najmanj dveh minut. uspešnosti sistema za SRG s posnetki za učenje modelov moških Vsako snemanje smo pričeli z branjem teksta nekega članka, pri slovenskih govorcev, ki smo jih posneli prek štirih kanalov. DET čemer se je govorec lahko vsaj približno privadil na snemalne krivulje uspešnosti sistema so na vseh grafih za mikrofonske naprave. Po branju smo v pogovoru, ki smo ga snemali, govorca posnetke obarvane rdeče, za PSTN posnetke črno, za GSM pripravili na ustrezno temo, ki mu je blizu. Pri tem je bil na posnetke zeleno in za VoIP posnetke modro. posnetkih slišen tudi govor sogovornika, ki je vodil pogovor in Ucenje-MIK Test-PSTN/GSM/MIK/VoIP/slo/ZS snemanje. Na ta način smo skušali zagotoviti čim bolj naraven in 80 sproščen način govora. V nadaljevanju smo posneli še spontani govor, ki je v obliki monologa o določeni temi trajal prav tako 2 minuti. Izkazalo se je, da govorcu močno olajšamo spontani govor 60 v obliki monologa, če se le ta smiselno in tematsko navezuje na pričeti pogovor v prejšnjem delu snemanja, saj se ljudje praviloma 40 počutijo nelagodno, ko morajo pred določeno osebo več časa )%( nepripravljeni govoriti o poljubni temi. RR 20 F Določeno težavo pri snemanju je predstavljala nesproščenost govorcev pri spontanem govoru. Izkazalo se je, da za določene ljudi 10 predstavlja nelagodje, če jih na snemanje vnaprej ne pripravimo. VoIP PSTN 5 Priprava je običajno obsegala obrazložitev postopka in namena MIK snemanja. Nekateri govorci so želeli, da jim zagotovimo GSM 2 anonimnost oziroma zagotovilo, da posnetki ne bodo zlorabljeni ali 1 javno objavljeni. 1 2 5 10 20 40 60 80 FAR (%) 4. EKSPERIMENT Slika 3. DET krivule učenja modelov z MIK posnetki. Meritev uspešnosti sistema za SRG [9] smo izvajali v več korakih: Ucenje-GSM Test-PSTN/GSM/MIK/VoIP/slov/ZS izbor posnetkov, generiranje modela ozadja, učenje sistema, 80 testiranje in analiza rezultatov. Najprej smo izbrali dve skupini posnetkov. Prva skupina posnetkov je bila namenjena za 60 generiranjem modela ozadja, druga skupina pa je bila razdeljena na podskupino za učenje in podskupino za testiranje sistema. 40 Pomembno je, da za model ozadja izberemo posnetke iz iste ) VoIP populacije, kot je zastopana v posnetkih za testiranje in učenje. Če %( R so za model ozadja uporabljeni posnetki govorcev, ki hkrati R 20 F MIK govorijo tudi na posnetkih za testiranje in učenje, govorimo o 10 testiranju v zaprtem podatkovnem setu. Model ozadja je pri vseh GSM meritvah enak in je zgrajen iz dveh skupin posnetkov. Prva skupina 5 vsebuje 66 posnetkov prek mikrofona, druga skupina pa 124 PSTN posnetkov prek mobilne telefonije. Za izgradnjo slovenskega 2 modela ozadja smo morali zaradi omejene količine slovenskih 1 govornih posnetkov vključiti tudi tiste posnetke, na katerih 1 2 5 10 20 40 60 80 FAR (%) govorijo iste osebe kot na posnetkih za učenje in testiranje. Obe skupini posnetkov za model ozadja sta bil sestavljeni iz posnetkov Slika 4. DET krivulje učenja modelov z GSM posnetki. pogovora in spontanega govora. Skupini mikrofonskih posnetkov smo torej dodali še 30 posnetkov moških govorcev iz javnih TV oddaj in parlamenta. Skupini posnetkov prek mobilne telefonije pa 21 Ucenje-TEL Test-PSTN/GSM/MIK/VoIP/slov/ZS okoli 20%. Testni posnetki preko GSM pa prinesejo napako EER 80 blizu 40% . 60 Tudi v primeru učenja z VOIP posnetki (slika 6) so rezultati najboljši v primeru, ko so učni in testni posnetki pridobljeni na enak 40 način. Opazimo pa lahko občutno poslabšanje napake pri posnetkih ) pridobljenih z mikrofonom in preko GSM telefonije z EER okoli (%RR 20 40%. Pri posnetkih pridobljenih s PSTN telefonijo znaša EER okoli F GSM 50%, kar pa je enakovredno naključnemu odločanju. 10 PSTN VoIP Iz vseh meritev v mešanih okoliščinah lahko ugotovimo, da se 5 MIK sistem za SRG pričakovano najbolje obnaša s posnetki, 2 pridobljenimi v istih učnih in testnih razmerah. Vrednost EER za 1 1 2 5 10 20 40 60 80 učne in testne posnetke pridobljene v enakih okoliščinah je FAR (%) razmeroma majhna, pod 5%, razen pri GSM, kjer je okoli 10%, pri Slika 5. DET krivulje učenja modelov z PSTN posnetki. čemer je pri VoIP celo pod 1%, tako da modra krivulja niti ni več vidna na grafu. To pripisujemo razmeroma majhni testni zbirki Ucenje-VoIP Test-PSTN/GSM/MIK/VoIP/slov/ZS 80 posnetkov. Gre namreč za testiranje v zaprtem podatkovnem setu. 60 PSTN 6. ZAKLJUČEK 40 Prdstavili smo problematiko razpoznavanja oz. verifikacije ) govorcev v forenzične namene. Poudarek je bil na problematika (%RR 20 pridobivanja posnetkov pod »stvarnimi pogoji«. V zvezi s tem smo F GSM proučevali vpliv različnih načinov zajemanja govornega signala na 10 rezultate prepoznavanja SRG. Za potrebe eksperimenta je bila 5 MIK posneta posebna govorna zbirka slovenskih govorcev, ki jo je 2 smiselno dograjevati z novimi glasovi. 1 1 2 5 10 20 40 60 80 Izkazalo se je, da so sistemi za SRG še vedno precej občutljivi na FAR (%) vplive prenosnega kanala. Največji izzivi so v primerih, ko Slika 6. DET krivulje učenja modelov z VoIP posnetki. izvajamo učenje sistema na podatkih, ki so pridobljeni preko ene vrste telefonije, testiranje sistema pa se izvaja na podatkih, ki so posneti preko druge vrste telefonije oziroma neposredno preko 5. REZULTATI mikrofona. Izkazalo se je, da so rezultati v primerih mešanih Iz rezultatov meritev uspešnosti sistema za SRG v mešanih pogojev znatno slabši od rezultatov pridobljenih v enakih pogojih. okoliščinah lahko ugotovimo, da se sistem za SRG pričakovano najbolje obnaša s posnetki, pridobljenimi v istih razmerah tako za učenje kot testiranje. 7. LITERATURA IN VIRI [1] Ortega-Garcia, J., Gonzalez-Roidriguez, J., Marrero-Aguiar, FRR (angl. False Rejection Rate) je verjetnost, da sistem za SRG V., 2000. AHUMADA: A large speech corpus in Spanish for ne zazna govorca na posnetku, kjer je govorec prisoten. Govorimo o deležu napačno zavrnjenih govorcev. speaker characterization and identification, Speech Communication 31, str. 255-264. FAR (angl. False Acceptance Rate) predstavlja verjetnost, da bo sistem za SRG napačno zaznal govorca, ki ni prisoten v posnetku, [2] Šef, T., Baucon, P., 2007. Sodno izvedenstvo in razpoznavanje (identifikacija) govorcev v kazenskem ki ga sistem analizira. Pri identifikaciji bo sistem identificiral govorca, ki ni prisoten v tesni množici, pri verifikaciji pa bo sistem postopku, Pravosodni bilten, 2/2007. napačno potrdil istovetnost neavtentičnega posnetka. [3] Hollien, H.,2002. Forensic Voice Identification, Academic Press. EER (angl. Equal Error Rate) predstavlja točko, kjer je verjetnost za napačno sprejetje in napačno zavrnitev enaka; torej je odločitev [4] Rose, P., 2005. Technical forensic speaker recognition: enaka naključnemu odločanju. Nižja kot je vrednost EER, boljši je evaluation, types and testing of evidence, Computer Speech sistem. and Language. Pri učenju z mikrofonskimi posnetki (slika 3) dosega EER pod 5 % [5] Alexander, A., 2005. Forensic Automatic Speaker za mikrofonske testne posnetke. Obnašanje sistema s PSTN Recognition Using Bayesian Interpretation and Statistical posnetki je nekoliko slabše, najslabše pa se sistem obnaša z GSM Compensation for Mismatched Conditions, doktorska in VoIP posnetki, kjer je EER okoli 20 %. disertacija, Lausanne, EPFL. Pri učenju s posnetki pridobljenimi preko mobilne telefonije GSM [6] Rose, P., 2002. Forensic speaker Identification, Taylor & (slika 4) dosežemo EER okoli 10 % v primeru GSM tesnih Francis. posnetkov, nekaj nad 10 % v primeru PSTN posnetkov, 15 % v [7] Jessen, M., 2007. Some Experiences from Tests of an primeru uporabe mikrofona in nad 20 % pri uporabi testnih Automatic Speaker Recognition System under Forensic posnetkov preko VOIP. Conditions, Bundeskriminalamt, EAFS. V primeru učenja sistema za SRG s posnetki preko telefonije PSTN [8] Blatnik, R., 2012. Vpliv kakovosti govora v telefoniji na (slika 5) najboljši rezultat zasledimo pri testnih posnetkih, ki so samodejno razpoznavanje govorca, magistrska naloga. posneti v enakih razmerah kot za učenje. Občutno slabše pa se sistem obnaša pri VoIP in mikrofonskih [9] http://www.persay.com/pdf/SPID_V6_DataSheet.pdf testnih posnetkih z EER 22 JSI Sound – platforma za enostavno klasifikacijo zvočnih posnetkov: Demonstracija na zvokih živali Borut Budna, Martin Gjoreski, Anton Gradišek, Matjaž Gams Odsek za inteligentne sisteme, Institut “Jožef Stefan” Jamova cesta 39, SI-1000 Ljubljana anton.gradisek@ijs.si POVZETEK so si izrazito različni med seboj, tako v časovni kot v frekvenčni Predstavljamo orodje JSI Sound, ki je namenjeno enostavni domeni. klasifikaciji zvočnih posnetkov. Implementirano je v okolju Orange, ki je odprtokodno orodje za strojno učenje in vizualizacijo podatkov za strokovnjake in začetnike. JSI Sound je bil razvit v skladu s paradigmo »strojno učenje kot storitev«, saj omogoča enostavno testiranje klasifikacijskih modelov na različnih bazah podatkov zvočnih posnetkov, v mislih imamo predvsem različne biozvoke. S tem je primeren tako za ljubitelje s področja bioakustike brez naprednega znanja s področja strojnega učenja, ki lahko JSI Sound uporabijo kot enostavno klasifikacijsko orodje, kot tudi za strokovnjake, ki ga lahko uporabijo za enostavno testiranje modelov kot prvi korak pri izdelavi specializiranih klasifikacijskih aplikacij. Vhodne podatke za JSI Sound predstavlja serija označenih posnetkov. Uporabnik v Slika 1: Spektrogram oglašanja ptice Sylvia communis (rjava seriji korakov s pomočjo grafičnega vmesnika izbere način penica). Iz spektrograma je razvidno izrazito strukturirano filtriranja, segmentacije in postopek določitev značilk. Na podlagi oglašanje, tako v časovni kot tudi v frekvenčni domeni. teh značilk orodje zgradi serijo klasifikacijskih modelov in jih testira. Tu predstavimo testiranje sistema na treh serijah podatkov – na brenčanju čmrljev ter oglašanju ptic in žab. Ključne besede Živalsko oglašanje, strojno učenje, Orange, klasifikacija 1. UVOD Metode umetne inteligence in strojnega učenja so dolgo temeljile na analizah strogo strukturiranih podatkov, v novejšem času pa se vedno bolj ukvarjajo z direktnimi podatki iz realnega sveta, kot so Slika 2: Spektrogram klicev naključnega netopirja. Gre za video in avdio posnetki. V tem prispevku se osredotočimo na kratke eholokacijske klice v ultrazvočnem območju. analizo in klasifikacijo zvokov, ki jih proizvajajo živali. Naloga je pomembna v zoologiji, denimo v študijah biotske raznolikosti. Čeprav je mnoge živalske vrste enostavno prepoznati na podlagi videza, to ni vedno možno – bodisi zaradi življenjskega sloga (mnoge ptice se denimo skrivajo v grmovju ali v trsju) bodisi zaradi tega, ker so si osebki več vrst tako podobni, da jih lahko ločimo šele ob podrobnem morfološkem pregledu – to pa zahteva, da osebek ujamemo. Klasičen primer so penice, skupina ptic, ki so si na videz precej podobne, se pa vsaka vrsta izrazito drugače oglaša, kar lahko uporabimo kot osnovo za prepoznavanje. Drug primer so netopirji – med letom ponoči je vizualno prepoznavanje izredno zahtevno ali celo nemogoče, lahko pa jih prepoznavamo na podlagi oglašanja – netopirji se oglašajo v ultrazvočnem Slika 3: Spektrogram brenčanja čmrlja Bombus griseocollis, območju, zvoke pa uporabljajo za eholokacijo in za delavka. Brenčanje je v časovni domeni precej neodvisno, v sporazumevanje. Še en primer so žuželke na travniku, te frekvenčni domeni pa je razvidna struktura osnovne proizvajajo različne vrste zvokov, kot so klici za sporazumevanje frekvence in višje harmonskih frekvenc. ali zvok brenčanja med letom. Slike 1-3 prikazujejo primere spektrogramov za oglašanje ptice, netopirja in čmrlja. Vidimo, da Problem klasifikacije živalskih vrst z metodami strojnega učenja na podlagi oglašanja ni nov – v literaturi zadnjih let najdemo 23 primere za različne skupine živali in za različne pristope. Gradišek Za ekstrakcijo značilk uporabimo odprtokodne knjižnice, trenutno et al. [1] so uporabili kombinacijo metod strojnega učenja za uporabljamo OpenSmile [10] in pyAudioAnalysis [6]. Te prepoznavanje nestrukturiranega brenčanja čmrljev, kot najboljša knjižnice podpirajo veliko število značilk (ki so rezultati metoda se je izkazal naključni gozd. Ganchev in Potamitis [2] sta kompleksnih matematičnih operacij na vhodnih posnetkih), tako v se ukvarjala z oglašanjem žuželk (črički, škržati, prave cvrčalke), časovni kot tudi v frekvenčni domeni. Značilke temeljijo na uporabila sta kombinacijo probabilističnih nevronskih mrež in operacijah, kot so mel-frekvenčni kepstralni koeficienti (MFCC), gaussovskih modelov. Dosegla sta 90 % klasifikacijsko točnost na koeficientih percepcijskega linearnega napovedovanja (PLP) in bazi 307 vrst. Stowell in Plumbey [3] sta uporabila koeficientih Chroma [10]. Ta pristop se je že izkazal koristnega nenadzorovano učenje za prepoznavanje ptic, različne metode za pri analizi človeškega govora [11]. prepoznavanje ptic so preizkušali tudi Cheng et al. [4]. Kot smo videli na primerih spektrogramov, ima oglašanje vsake od skupin živali svoje posebnosti, zato se ni enostavno odločiti, na kakšen način bomo pristopili h klasifikacijskemu problemu – kakšne značilke izbrati in katere algoritme uporabiti. Odločitev je še težja za strokovnjake s področja bioakustike, ki bi si želeli delujoče klasifikacijske aplikacije, nimajo pa obširnega znanja s področja strojnega učenja. Naša rešitev JSI Sound [5] izhaja iz paradigme »strojno učenje kot storitev«. Izdelali smo orodje, s katerim lahko uporabnik s pomočjo grafičnega vmesnika preizkusi različne pristope pri gradnji klasifikacijskih modelov. Po eni strani je to lahko že dovolj za enostavne klasifikacije, po drugi strani pa predstavlja dobro osnovo za gradnjo robustnih specializiranih aplikacij. Naše orodje se razlikuje od obstoječih rešitev, ki jih lahko razdelimo predvsem na orodja v obliki knjižnic ali modulov (npr. pyAudioAnalysis [6]) ter orodja z grafičnim vmesnikom (kot je denimo Audacity [7]). Primerjava nekaterih funkcij za vsako od orodij je prikazana v Tabeli 1. Tabela 1: Primerjava funkcij treh različnih orodij za obdelavo zvoka Audacity pyAudioAnalysis JSI Sound Spektrogram da da da Slika 4: Shema metode, ki jo uporablja orodje JSI Sound Na podlagi značilk Orange zgradi odločitvene modele, kot so Filtriranje da ne da naključni gozd, SVM, naivni Bayes in druge. Za evaluacijo Segmentacija da (ročna) da da posameznega modela se podatke razdeli na učno množico, na kateri se trenira modele, in na testno množico, na kateri se te Ekstrakcija ne da da modele nato testira. Pri tem JSI Sound skrbi, da so vsi segmenti, značilk ki pripadajo istemu začetnemu posnetku, vedno ali v učni ali v testni množici. Poleg tega omogoča gradnjo modelov na nivoju Klasifikacija ne da da posameznega segmenta ali pa na nivoju celotnega posnetka, z GUI da ne da uporabo kombinacije napovedi za vsakega od segmentov posebej. Orodje JSI sound je razvito v okolju Orange [8,9], s tem je prosto 3. IMPLEMENTACIJA V OKOLJU dostopno. V prispevku opišemo splošno delovanje metode, ORANGE implementacijo v okolju Orange in rezultate testiranja na treh skupinah posnetkov živalskih zvokov. Po namestitvi je vtičnik JSI Sound dostopen v klasičnem seznamu vtičnikov okolja Orange. Uporabnik naloži bazo posnetkov, nato s seznama izbere ustrezne filtre ter parametre za segmentacijo posnetkov (dolžina in prekrivanje okna). Zatem uporabnik izbere 2. METODA knjižnice za ekstrakcijo značilk. Splošna metoda je prikazana na Sliki 4. Sestavljena je iz petih V naslednjem koraku uporabnik s pomočjo orodij v okolju Orange korakov: vnos zvočnih posnetkov, predprocesiranje, ekstrakcija značilk, gradnja modelov strojnega učenja ter evaluacija modelov. gradi modele ter jih evalvira. Primer uporabe JSI Sound je prikazan na Sliki 5. Vhodne podatke predstavlja skupina označenih posnetkov, zaželeno je, da so bili vsi pridobljeni z enako opremo in z enako frekvenco vzorčenja. Predprocesiranje zajema uporabo filtrov za odstranitev šuma. Izberemo lahko med petimi filtri; FIR, Butterworth, Čebišev, Eliptični in Besselov filter. V tem koraku izvedemo tudi segmentacijo na posnetke izbrane dolžine. Uporabimo lahko fiksno ali drseče okno. Hkrati lahko zavržemo segmente, v katerih ni dovolj informacije. 24 Tabela 3: Klasifikacijski rezultati v metriki AUC z 10-kratnim prečnim preverjanjem za vsakega od modelov za vse tri skupine živali Ptice Žabe Čmrlji LR 94 100 80 Slika 5: Primer uporabe orodja JSI Sound v okolju Orange. kNN 92 99 75 Vtičniki, ki so pobarvani modro, so bili razviti za JSI Sound, RF 94 100 81 ostali so standardni vtičniki okolja Orange. Po opisu implementacije in sheme uporabe lahko uporabnik vidi, NB 89 99 74 da je orodje enostavno za uporabo, poleg tega pa mu omogoča AdaBoost 85 93 67 hitro in natančno gradnjo klasifikacijskih modelov, katerih natančnost je primerljiva z rezultati prikazanimi v Tabeli 3. Rezultati za ptice so odlični in bodo osnova za izdelavo namenske aplikacije. Visoka klasifikacijska točnost za žabe je verjetnost 4. EKSPERIMENTI posledica majhnega števila posnetkov v vsakem od razredov, za Orodje JSI Sound smo testirali na treh različnih skupinah izboljšanje zanesljivosti bo potrebnih več posnetkov. Rezultati za živalskih zvokov, na posnetkih oglašanja slovenskih žab in ptic iz čmrlje so blizu tistim, ki smo jih dobili na bazi posnetkov družine penic ter na posnetkih brenčanja čmrljev, gre za vrste iz slovenskih vrst [1]. zvezne države Kolorado v ZDA. Število razredov, posnetkov in segmentov za vsako od skupin prikazuje Tabela 2. Tabela 2: Struktura podatkov za vsako od skupin živali 5. ZAKLJUČEK Predstavljamo orodje JSI Sound, ki je bilo razvito z namenom Ptice Žabe Čmrlji olajšati naloge s področja strojnega učenja za uporabnike, ki Št. razredov 6 13 9 nimajo naprednih izkušenj s tega področja ali s področja obdelave zvočnih posnetkov. Orodje JSI Sound je implementirano v okolju Št. posnetkov 81 39 51 Orange in vsebuje pet metod filtriranja, dve metodi segmentacije posnetkov ter dve obsežni knjižnici za določanje značilk, tako v Št. segmentov 5536 4447 3854 časovni kot tudi v frekvenčni domeni. Vse te funkcije so dostopne kot vtičniki za Orange, do njih pa dostopamo prek grafičnega vmesnika. Eksperiment je sestavljen iz treh korakov: strojnega učenja na nivoju segmentov, ekstrakcije značilk na nivoju posnetkov in Sistem smo testirali na treh setih posnetkov živalskih zvokov – na strojnega učenja na nivoju posnetkov. Motivacija za ta pristop je posnetkih brenčanja čmrljev ter oglašanju ptic iz družine penic ter žab. Rezultati klasifikacijskih modelov so podobni tistim, ki smo dvojna: ker vemo, da posamezni segmenti istega posnetka pripadajo istemu razredu, nam kombinacija informacij o več jih na istih ali podobnih setih dobili v prejšnjih študijah, kar kaže segmentih lahko pove več kot le informacija o posameznem na primernost orodja JSI Sound za tovrstne naloge. V teku je posnetku. Več metod strojnega učenja pa kombiniramo zato, ker raziskovanje primernosti metode za analizo drugih tipov zvokov, lahko različne metode delujejo različno dobro na posameznih denimo za analizo govora in zvokov človeškega telesa, kot sta strukturah v podatkih. bitje srca in dihanje. Za strojno učenje na podlagi posnetkov smo uporabili sledeče metode: Logistična regresija (LR), Naivni Bayes (NB), metoda najbližjega soseda (kNN), naključni gozd (RandomForest, RF) in 6. ZAHVALA AdaBoost. Vhodni podatek za vsako od metod je vektor značilk Zahvaljujemo se dr. Tomiji Trilarju za posnetke ptic in žab, Primožu Presetniku za posnetek netopirja ter prof. Candace Galen za vsakega od segmentov, izhodni podatek pa so verjetnosti za in dr. Nicole Miller-Struttmann za posnetke čmrljev. vsakega od klasifikacijskih razredov. V koraku ekstrakcije značilk na podlagi posnetkov združimo napovedi modelov iz predhodnega koraka, uporabimo 7. VIRI maksimalno, minimalno in povprečno vrednost napovedi vsakega [1] Gradišek, Anton, Gašper Slapničar, Jure Šorn, Mitja Luštrek, od modelov. Matjaž Gams, and Janez Grad. 2017. 'Predicting species V koraku strojnega učenja na osnovi celotnih posnetkov na identity of bumblebees through analysis of flight buzzing podlagi napovedi posameznih modelov na nivoju segmentov sounds', Bioacoustics, 26: 63-76. izdelamo meta-klasifikatorje. Te preverimo s desetkratnim [2] Ganchev, Todor, and Ilyas Potamitis. 2007. 'Automatic prečnim preverjanjem na testni množici. Rezultati, v metriki acoustic identification of singing insects', Bioacoustics, 16: ploščine pod krivuljo (area under curve, AUC), so predstavljeni v 281-328. Tabeli 3. [3] Stowell, Dan, and Mark D Plumbley. 2014. 'Automatic large- scale classification of bird sounds is strongly improved by unsupervised feature learning', PeerJ, 2: e488. 25 [4] Cheng, Jinkui, Bengui Xie, Congtian Lin, and Liqiang Ji. Principles of Data Mining and Knowledge Discovery, 537- 2012. 'A comparative study in birds: call-type-independent 39. Springer. species and individual recognition using four machine- [9] Demšar, Janez, Tomaz Curk, Aleš Erjavec, Črt Gorup, learning methods and two acoustic features', Bioacoustics, Tomaž Hočevar, Mitar Milutinovič, Martin Možina, Matija 21: 157-71. Polajnar, Marko Toplak, and Anže Starič. 2013. 'Orange: [5] Budna, Borut. 2017. 'Platform for audio clips classification', data mining toolbox in Python', Journal of Machine Learning University of Ljubljana, Faculty of Computer and Research, 14: 2349-53. Information Science. [10] Eyben, Florian, Martin Wöllmer, and Björn Schuller. 2010. [6] https://github.com/tyiannak/pyAudioAnalysis "Opensmile: the munich versatile and fast open-source audio [7] http://www.audacityteam.org/ feature extractor." In Proceedings of the 18th ACM international conference on Multimedia, 1459-62. ACM [8] Demšar, Janez, Blaž Zupan, Gregor Leban, and Tomaz Curk. 2004. "Orange: From experimental machine learning to [11] Martin Gjoreski, Hristijan, Gjoreski, and Andrea Kulakov. interactive data mining." In European Conference on 2014. Machine learning approach for emotion recognition in speech. Informatica, vol. 38, no. 4, pp. 377-384. 26 Bat Classification using Deep Neural Network Jani Bizjak Anton Gradišek Luka Stepančič Primož Presetnik Department of Intelligent Department of Intelligent Department of Intelligent Center za Kartografijo Favne Systems, Systems, Systems, in Flore Jožef Stefan Institute Jožef Stefan Institute Jožef Stefan Institute Klunova 3 Jamova cesta 39 Jamova cesta 39 Jamova cesta 39 1000 Ljubljana, Slovenia 1000 Ljubljana, Slovenia 1000 Ljubljana, Slovenia 1000 Ljubljana, Slovenia primoz.presetnik@ckff.si +386 1 477 3147 +386 1 477 3147 +386 1 477 3147 jani.bizjak@ijs.si anton.gradisek@ijs.si luka.stepancic@ijs.si ABSTRACT between two species of bats. Visual recognition can often be a We present a deep neural network approach for bat species difficult task when dealing with lowlight conditions, which are classification by echolocation and social calls. First the data is usually present when bats are normally active. On the other hand, gathered on two separate locations using special high frequency classification based on bat calls is a more promising approach. ultrasound recorder. The data is then preprocessed in order to be Experts can classify most of Slovenian bat species based on the usable in deep neural network architecture. Deep architecture used audio recordings. However, automatic classification methods are is discussed and experimental results for classification of two desired since they allow us to process large amounts of data from species are presented. The last part of the paper focuses on future recording stations. work that could improve results. In last years deep neural networks (DNN) proved to perform very well in classification of real world signals such as sounds or images General Terms [2]. In this paper we present an advanced deep learning architecture Algorithms, Design, Experimentation that can be used for bat species classification based on sound. Keywords bats, deep neural network, convolutional neural networks, Pipistrellus pygmaeus, Barbastella barbastellus 1. INTRODUCTION Bats are second largest order of mammals, representing 20% of all mammal species worldwide. Bats live in most of the world except extremely cold regions. There are 30 species of bats classified in Slovenia out of more than 1200 found around the world. Bats in Slovenia are all insectivores – meaning they feed on insects. Figure 1: Two species of bats found in Slovenia that we try to They are mostly active during the dusk and in general have poorly classify. developed vision. For navigation and hunting however they use 2. Data special organ that works similarly to sonar. They emit ultrasonic Data consist of two datasets from two different location Dolenja sounds, between 10 and 120 kHz and are precise enough to detect Vas; around 2000 recordings, 8 GB in size, and Kozina which less than a 0.01 mm wide obstacle [1]. contains 8.900 recordings and is 28 GB in size. They live in colonies and also have ultrasonic social calls in order The data were recorded using a specialized high frequency recorder to communicate with each other. The sounds they produce differs SM4BAT FS&ZC. Recorder has a recording range of 16 Hz to 150 from specie to specie and can be used for classification. kHz and recording frequency of 500 kHz. We used band pass filter Bats perform vital ecological roles of pollinating flowers, they between 10 kHz to 110 kHz in order to filter out most of the noise consume insect pests and their excrements (guano) are very good and to lower amount of empty recordings (in general bats do not fertilizers. It is thus important for humans to do what they can to produce noise that is lower than 15 kHz or higher than 100 kHz at help them prosper. In this paper we focus on two species found in the same time there are not many animals that produce sounds in Slovenia, Barbastella barbatellus and Pipistrellus pygmaeus such frequency range). [Figure 1]. The recorder automatically records a sound, couple of seconds Modern ICT solutions nowadays allow us to better understand and earlier and couple of seconds after certain frequency threshold is study these animals by using automated systems. Since bats use reached e.g. when sound over 15kHz is detected. The recording is such a unique form of communication they can be easily detected then manually labeled by the expert, who normally looks at the using sound compared to other methods (image), however sound spectrogram in order to determine number and species of bats in the data is full with noise and it might not be an easy task to differ recording. In total the dataset contains 22 different species. Majority of cases consists of one specie per recording although 27 some contain multiple species. In some recordings expert was 3.1 Deep neural network unable to determine exact specie so a specie family is used as a Artificial neural networks have existed since 1954 when Farley and label. In total there are 37 labels in this dataset. Weasley first implemented a simple neural network on a computer. 2.1 Data preprocessing Because of the computation complexity they were not used widely until the late 2000’s, when computers with new architectural The information in wave form is concentrated in certain frequency design, initially optimized for graphic cards to allow components which are impossible to detect in wave form format parallelization, became capable enough to run several layers of that is why Furrier transform is performed to transform data into neurons – a deep architecture. Architecture is considered deep if it time-frequency representation – spectrogram. Spectrogram format contains at least 4 layers of neurons. also gives us a better look into frequency distribution in a signal (it is easier for human to interpret a visual information). The frequency In 2012, AlexNet [3] architecture was proposed for image resolution of spectrogram used was 256. recognition. The network achieved more than a 10 % increase in accuracy compared to the second best method and made DNNs one A representative spectrogram is shown in Figure 2. It is easy to of the most used ML methods today. The architecture uses distinguish 3 types of signals. Below 20 kHz there is noise, which convolutional layers to generalize input image and combine can be discarded. At higher frequencies, there are two types of previously learned features into more complex high level features. patterns. A social call, which spans over a higher range of By the rule of thumb, Convolutional Neural Networks (CNN) can frequencies, and an echolocation call, that is used for navigation. be used and often perform good, especially when the task can be Figure 2: Spectrogram of one recording. In the figure 3 distinct patterns emerge. Everything below 20 kHz is noise, in the range of 20kHz and 100 kHz there are two type of sounds, social calls and echolocation calls. Due to the high sampling frequency (500kHz) of the recording presented in such a way that a person can use their bare eyes to device it is not possible to feed whole signal to the neural network. classify the objects. Since spectrograms are analyzed by humans by We split the recording using a sliding window of size 2048 samples. looking at them and discovering patterns in the time-frequency To reduce the size of the file we also removed frequencies below representation it is likely that CNN will perform better than other 15 kHz and above 80 kHz, which did not contain any relevant architectures on the same domain. information for classification of the two bat species that we We loosely based our architecture on several state-of-the-art attempted to identify. architectures [3,4,6] and fine tune it to achieve best results for bat When dealing with natural signals such as sound or image, it is best domain. In our architecture we used 3 convolutional layers, each to feed the neural networks with raw signal and allow for their consisting of 34 filters sized 7x7, 5x6, and 3x3 on each consecutive abstraction power to generalize information out of the data. layer. After the second and third layer, we use MaxPooling layer Spectrogram can be represented to the neural network as an image which uses filter size 2x4 and 2x2. Max pooling is used to reduce - that is why we do not attempt to manually extract any more size of the image (data). Filter slides over the image and features but feed the windows extracted to the network. concatenates all values under the filter into one. Different approaches can be used e.g. average, min, or max [4]. 3. DEEP ARCHITECTURE Deep neural networks are becoming increasingly popular in the last years. They perform especially well on natural signals and are dominant on the domains of image recognition, voice recognition and natural language recognition. In our experiments we tested several deep architectures that we present in the following sections. 28 After the fifth layer, 3 fully-connected layers follow. In order to reduce over-fitting, a small number of neurons is used 8, 4, and 2 on layers 6, 7, 8. We use rectified linear activation function (ReLU) [Figure 3] which in general gives best performance [5]. Figure 4: Network Architecture, uses 3 convolutional layers, 2 MaxPooling and 3 fully connected layers in the end. Despite large amount of data in total there are only 128 recordings of the selected species. We split the recordings into 793 half second windows using sliding window technique. The amount of data available for training is extremely small for deep learning. Even Figure 3: Rectifier Linear Function gives good worse is that a lot of windows from this data only contain noise and performance and is extremely easy to calculate no bat calls. on GPU where float operations are computational expensive (example sigmoid) We split the data 75% for training and 25% for testing and despite all achieved average 8% improvement compared to the majority In the last layer we use SoftMax regression for classification. The class, which indicates that the networks were able to learn whole architecture is presented in Figure 4. something from the data. In order to speed up the learning, instead of using standard gradient descend (SGD) for adjusting neuron weights we use RMSProp [4] 5. FUTURE WORK which supports batched learning - allows for parallelization, adjust Our initial experiments showed that deep neural networks have gradient weight for each parameter separately and uses adaptive potential in bat classification based on sound. Currently the main normalization with decay parameter β [Equation 1]. problem is data segmentation and its labels. The one label per 𝑣 = 𝛽𝑣 + (1 − 𝛽)(∆𝑓)2 recording, which is then divided into multiple windows (with the same label) brings a lot of noise into the data. One solution would ∆𝑓 𝜃 = 𝜃 − 𝛼 be to use whole recording as one instance of a class, but because of √𝑣 + 𝜀 vast amount of data in one recording it is not currently feasible to Equation 1 RMSProp do it. There are two solutions for this problem: have a better pre- processing and only learn on correctly labeled segments or introduce a new architecture of deep neural networks that can learn Despite large amounts of data, we discovered our network was still on unlabeled datasets. over-fitting. In order to resolve this problem, we introduced dropout [6] in the last layer. During training phase, the dropout For our next step we have implemented a segmentation method that method will randomly remove connections from neurons in one uses power envelope to isolate bat calls from random noise or layer making the layer temporarily sparsely connected. The empty recordings as seen in Figure 5. Our initial tests show removed connections will be shuffled after each iteration preventing co-adaptations of neurons during learning. In the testing phase the model is again fully connected (dropout only works during learning phase). We used dropout probability of 0.5. 4. EXPERIMENTS Recordings in the dataset last multiple seconds and contain several calls from one or multiple bats. Because of the large sampling frequency, it is not possible to input the whole recording into a neural network so we use windows. The problem with windows is that they are not labeled separately but have the same label as the original recording. This presents a problem with windows that do not contain any bat call. In order to alleviate this problem, we only focused on distinguishing between two species of bats Barbastella barbatellus and Pipistrellus pygmaeus. They were chosen because they had better noise to data ratio. Figure 5: Bat calls segmentation using power envelope promising results (around 2 times better precision over majority 29 class). However, it is likely that some of the time/sequential information is lost with this process. 7. REFERENCES In order to avoid this, the segmentation method can be used to first [1] Slovensko Društvo za Proučevanje in Varstvo Netopirjev, train the network to detect any bat sounds. When the network is http://www.sdpvn-drustvo.si/ (accessed 2017) proficient enough in the later task it can then be used to differentiate between different species. By dividing the problem into two sub [2] Schmidhuber, Jürgen. "Deep learning in neural networks: An problems the complexity/depth of the network can be lowered overview." Neural networks 61 (2015): 85-117.. which allows for faster learning time and mitigates overfitting to an [3] Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. extent. "Imagenet classification with deep convolutional neural networks." Advances in neural information processing 6. CONCLUSION systems. 2012. In this paper we presented problem of bat classification based on [4] Bizjak, Jani. Analiza signalov EKG z globokimi nevronskimi ultrasound recordings. We recorded an extensive database of bat mrežami: magistrsko delo. Diss. J. Bizjak, 2015. recordings and presented a deep architecture used for species classification. In last part we presented initial results and proposed [5] Maas, Andrew L., Awni Y. Hannun, and Andrew Y. Ng. methods in future work to improve current model. "Rectifier nonlinearities improve neural network acoustic models." Proc. ICML. Vol. 30. No. 1. 2013. Despite the initial poor results, which we think are the result of noisy data, we have shown that deep neural networks can be used [6] Srivastava, Nitish, et al. "Dropout: a simple way to prevent for animal sound classification – more precisely bats. We believe neural networks from overfitting." Journal of machine that with more data and better segmentation greater improvements learning research 15.1 (2014): 1929-1958. in accuracy can be achieved. Additionally, semi-supervised learning can be used to train the model on vast amount of unlabeled data. 30 Globoke nevronske mreže in matrična faktorizacija Gašper Petelin Igor Kononenko Univerza v Ljubljani, FRI Univerza v Ljubljani, FRI Večna pot 113 Večna pot 113 Ljubljana Ljubljana gp6279@student.uni-lj.si igor.kononenko@fri.uni-lj.si POVZETEK Slika 1 prikazuje predlagano inicializacijo za ciljno mrežo z Predlagana je nova inicializacija globokih nevronskih mrež, ki dvema skritima nivojema. Prvi korak je faktorizacija matrike 𝑋 v pred začetkom učenja nevronske mreže nastavi uteži in dve novi matriki 𝑊1 in 𝑊2. Ko izračunamo abstraktno pristranske vrednosti tako, da usmeri nevronsko mrežo proti predstavitev podatkov za oba skrita nivoja, zgradimo tri rešitvi, ki pri optimizaciji hitreje konvergira k lokalnemu enonivojske mreže, kjer prva mreža preslika originalno matriko 𝑋 minimumu. Predlagana inicializacija je od običajne inicializacije v prvo zgoščeno vrednost 𝑊1, druga mreža slika iz 𝑊2, zadnja precej počasnejša, vendar lahko pri globokih nevronskih mrežah mreža pa iz 𝑊2 v končno matriko 𝑌. precej pohitri skupni čas učenja. Za učenje enonivojskih mrež lahko uporabimo različne funkcije Ključne besede napak z različnimi aktivacijskimi funkcijami. Pri učenju enonivojskih mrež, je pomemben tudi izbor števila iteracij (angl. Globoke nevronske mreže, klasifikacijska točnost, regresija, epoch) in velikost učnega paketa (angl. batch), saj lahko pride do nenegativna matrična faktorizacija, analiza arhetipov, prevelikega prileganja, kar upočasni učenje ciljne mreže. inicializacija uteži . 1. UVOD Globoke nevronske mreže v zadnjih letih dosegajo precej dobre rezultate [6][7], vendar se z večanjem števila skritih nivojev pojavijo tudi nekatere težave, kot so izguba gradienta pri vzvratnem širjenju napake in čas, ki je potreben za učenje. Predlagana je nova metoda za nastavljanje začetnih uteži, ki temelji na inicializaciji s pomočjo matrične faktorizacije, kjer je cilj, da bi pospešili učenje in preprečili preveliko izgubo gradienta. 2. PREDLAGANA INICIALIZACIJA Ideja za predlagano inicializacijo delno izhaja iz inicializacij uteži z algoritmi za nenadzorovano učenje, kjer poskušamo vsak nivo nevronske mreže naučiti čim bolj abstraktno predstavitev podatkov, iz katerih lahko nato lažje pravilno napovemo vrednosti na izhodu. Pri predlaganem algoritmu te abstraktne predstavitve podatkov izračunamo s pomočjo matrične faktorizacije. Slika 1. Izgradnja in učenje enonivojskih mrež ter njihovo združevanje v ciljno nevronsko mrežo. Predpostavimo, da imamo globoko nevronsko mrežo z k skritimi nivoji, pri kateri bi se radi naučili preslikati matriko podatkov 𝑋 ∈ Zadnji korak inicializacije je, da uteži teh enonivojskih mrež ℝ𝑛×𝑚 v izhodne podatke 𝑌 ∈ ℝ𝑛×ℎ. Konstante n, m in h uporabimo za inicializacijo ciljne globoke nevronske mreže. Slika predstavljajo število učnih primerkov, dimenzijo učnih podatkov 1 prikazuje postopek, kjer enonivojske mreže združimo v ciljno in število nevronov na izhodu ciljne nevronske mreže. Originalno mrežo, ki jo še dodatno učimo. matriko 𝑋 najprej faktoriziramo v pare {𝑊𝑘 ∈ ℝ𝑛×𝑖, 𝐻𝑘 ∈ 3. REZULTATI MNIST ℝ𝑖×𝑚}, kjer rank i predstavlja število nevronov ciljne nevronske mreže na skritem nivoju Za testiranje inicializacije za klasifikacijske probleme je bila k. Število teh parov je zato enako številu nivojev v ciljni nevronski mreži. uporabljena podatkovna množica MNIST, ki vsebuje slike ročno Dobljene matrike 𝑊 v tem napisanih številk. Slika 2 prikazuje hitrost učenja ciljne nevronske primeru predstavlja abstrakten povzetek originalne matrike mreže inicializirane s predlagano inicializacijo. Razvidno je, da za oziroma aktivacije ciljne mreže na posameznem skritem nivoju. globoko nevronsko mrežo učenje poteka precej hitreje kot Matrik 𝐻 za inicializacijo ne potrebujemo, zato jih lahko inicializacija z naključno zavržemo. inicializacijo Xavier. Pri nevronskih mrežah z manj kot 5 skritimi nivoji, je ta Naslednji korak inicializacije je učenje k+1 enonivojskih mrež, inicializacija največkrat nepotrebna, saj med učenjem ne prihaja kjer se vsaka od mrež nauči preslikavo iz ene predstavitve do tako velikih izgub gradienta, kot je to pri globokih nevronskih podatkov 𝑊𝑘 v drugo predstavitev 𝑊𝑘+1. Posebnost so le prva mrežah. enonivojska mreža, ki slika iz matrike 𝑋 v 𝑊1 in zadnja enonivojske mreža, ki slika iz 𝑊 𝑘 v matriko izhodnih podatkov 𝑌. Za faktorizirane pare velja, da ni možno, da je dimenzija 𝑖 večja od dimenzije 𝑚 v matriki 𝑋 oziroma velja 𝑖1, 𝑖2, … , 𝑖𝑘 < 𝑚. 31 Slika 2. Primerjava klasifikacijske točnosti med predlagano Slika 4. Primerjava vpliva različnih tipov matrične inicializacijo in naključno inicializacijo Xavier na globoki faktorizacije na hitrost učenja ciljne nevronske mreže. nevronski mreži z arhitekturo nivojev (400, 300, 200, 100, 70, Ena izmed najboljših postopkov za pohitritev učenja in 50, 40, 30, 20, 15, 13, 10) in sigmoidno aktivacijsko funkcijo. preprečevanje nasičenosti aktivacijskih funkcij je uporaba funkcij, Ker lahko pri enonivojskih nevronskih mrežah pride do kot so ReLU ali katerokoli izmed mnogih variant te funkcije. prevelikega prileganja podatkom, kar slabo vpliva na Slika 5 primerja hitrost učenja med aktivacijsko funkcijo ReLU in klasifikacijsko točnost ciljne mreže, lahko že pri učenju teh mrež sigmoidno aktivacijsko funkcijo. dodamo regularizacijo, ki nekoliko prepreči preveliko prileganje ciljne mreže. Slika 3 prikazuje vpliv različnih stopenj regularizacije L2 na točnost ciljne mreže. Slika 5. Primerjava spreminjanje klasifikacijske točnosti pri uporabi aktivacijske funkcije ReLU in sigmoidne funkcije. Izbor aktivacijske funkcije zelo vpliva na hitrost učenja, kar je Slika 3. Vpliv različnih stopenj regularizacije L2 enonivojskih povsem pričakovano. Pri uporabi sigmoidne funkcije se mrež na klasifikacijsko točnost med učenjem ciljne nevronske predlagana inicializacija obnese precej dobro, saj se že po 100 mreže. iteracijah nauči precej dobre inicializacije, med tem ko se pri Iz stopenj regularizacije je razvidno, da lahko ob pravilni stopnji inicializaciji Xavier klasifikacijska točnost ne povzpne nad 20%. regularizacije enonivojskih mrež pospešimo hitrost učenja ciljne Pri uporabi funkcije ReLU z inicializacijo dosežemo hitrejše nevronske mreže. Če je regularizacija enonivojskih mrež učenje, vendar kar je najbolj zanimivo, že pred začetkom učenja prevelika, se te ne naučijo uporabne preslikave, kar zmanjša ciljne mreže dosega ta klasifikacijo nad 65%, med tem ko je hitrost učenja ciljne mreže. točnost pri naključni inicializaciji pred začetkom učenja ciljne Pri izračunu zgoščene matrike podatkov lahko uporabimo več mreže 10%. različnih metod za matrično faktorizacijo. Metode faktorizacije se Inicializacija s pomočjo matrične faktorizacije je precej počasen med seboj precej razlikujejo. Slika 4 prikazuje vpliv različnih postopek, saj večkrat faktoriziramo celotno matriko ter nato še metod matrične faktorizacije na klasifikacijsko točnost ciljne učimo enonivojske mreže. Slika 6 prikazuje inicializacijo, kjer nevronske mreže med učenjem. Uporabljeni so bili tipi NMF smo za faktorizacijo in učenje enonivojskih mrež uporabili le del (nenegativne matrična faktorizacija), AA (analiza arhetipov) [1], učne množice. SNMF (semi-nenegativna matrična faktorizacija), SVD (singularni razcep), PCA (analiza glavnih komponent) in FA Iz grafa je vidno, da lahko pri uporabi le dela učne množice za (faktorska analiza). Za primerjavo je zraven dodana še inicializacijo dobimo boljše rezultate kot pa pri uporabi celotne inicializacija Xavier, ki se v tem primeru odreže precej slabše kot množice, pri tem pa precej pridobimo še na hitrosti pri matrični druge inicializacije. Pri napovedovanju se najbolje odreže faktorizaciji in hitrosti učenja enonivojskih mrež. algoritem AA, ki najhitreje doseže najboljšo klasifikacijsko točnost. Najslabša sta običajno algoritma PCA in NMF. 32 Največja prednost predlagane inicializacije se je izkazala pri regularizaciji šal, saj je ciljna nevronska mreža pri šalah dosegla optimalno povprečno absolutno napako že po nekaj iteracijah, nato pa je prišlo do prevelikega prileganja. Slika 8 prikazuje MAE ciljne nevronske mreže pri različnih stopnjah regularizacije L1 enonivojskih nevronskih mrež. Iz napake MAE vidimo, da lahko že s pravilno nastavljenimi utežmi, ki jih dobimo z regularizacijo uteži enonivojskih mrež dosežemo podoben efekt, kot če bi regularizacijo uporabili med učenjem ciljne nevronske mreže. 5. ČAS, POTREBEN ZA INICIALIZACIJO Pri predlagani inicializaciji je potrebno najprej izvesti večje število matričnih faktorizacij, ki odvisno od časovne Slika 6. Primerjava klasifikacijske točnosti med učenjem ciljne mreže v odvisnosti od odstotka uporabljenih podatkov kompleksnosti potrebujejo precej časa. Ko končamo s faktorizacijo je potrebno še učenje enonivojskih mrež. V pri inicializaciji. določenih primerih je za inicializacijo potrebno celo več časa kot 4. REZULTATI JESTER JOKES pa za učenje ciljne nevronske mreže. Slika 9 primerja skupen čas, Inicializacija je bila testirana tudi za regresijski problem potreben za inicializacijo in učenje ciljne nevronske mreže, pri napovedovanja ocen šal, ki bi jih uporabniki dali šalam na spletni različnih odstotkih uporabljenih učnih podatkov za inicializacijo v strani Jester. odvisnosti od klasifikacijske točnosti ciljne mreže. Za primerjavo so bili uporabljeni podatki MNIST. Črtkana črta prikazuje čas, ki Slika 7 prikazuje hitrost učenja pri uporabi različnih tipov je potreben za inicializacijo in je odvisen od odstotka učne matrične faktorizacije. Najboljša je bila ponovno faktorizacija množice, uporabljene za inicializacijo. Pri uporabi le 30% AA, najslabša pa PCA, ki se sploh ni učila. podatkov se inicializacija izvede že po 150 sekundah, medtem ko inicializacija pri uporabi celotne množice potrebuje skoraj 500 sekund. Naključna inicializacija Xavier je na začetku boljša od predlagane inicializacije, saj že hitro po začetku učenja dosega 20% točnost, vendar je učenje precej počasno, zato jo druge nevronske mreže, ko so enkrat inicializirane hitro prehitijo po kriteriju klasifikacijske točnosti. Za merjenje časa je učenje nevronskih mrež potekalo z grafično kartico GTX 960M. Za faktorizacijo in učenje nevronskih mrež pa so bile uporabljene knjižnice sklearn, PyMF in Keras [2]. Slika 7. Primerjava vpliva različnih tipov matrične faktorizacije na hitrost učenja ciljne nevronske mreže. Slika 9. Primerjava časa, ki ga potrebuje predlagana inicializacija za izračun uteži v odvisnosti od klasifikacijske točnosti med učenjem. Matriki 𝑊 in 𝐻 sta pred začetkom faktorizacije inicializirani naključno. Prav tako so naključno inicializirane uteži enonivojskih mrež. Zaradi naključne inicializacije so lahko rezultati med posameznimi testi nekoliko različni, zato so vsi grafi dobljeni kot povprečje treh poganjanj algoritma. Isto velja za graf časov, ki so sestavljeni kot povprečje treh testov. Slika 8. Napaka MAE ciljne mreže pri uporabi različnih stopenj regularizacije L1 enonivojskih mrež. 33 6. ZAKLJUČEK internal covariate shift," in International Predstavljena in testirana je inicializacija, kjer začetne vrednosti Conference on Machine Learning, 2015. uteži izračunamo s pomočjo matrične faktorizacije. Metoda se dobro obnese pri podatkih, ki imajo precej veliko število atributov in za napovedovanje izhodnih vrednosti potrebujejo globoke [4] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio nevronske mreže, saj se mreža že v začetku nauči nekatere and P.-A. Manzagol, "Stacked denoising abstraktne koncepte, ki obstajajo v podatkih. autoencoders: Learning useful representations in a 7. PODOBNE INICIALIZACIJE deep network with a local denoising criterion," Skozi razvoj nevronskih mrež so te postajale vsakič bolj globoke, Journal of Machine Learning Research, vol. 11, da bi se lahko naučile čim bolj zapletenih povezav med podatki. Pri globljih mrežah pp. 3371-3408, 2010. se pri naključni inicializaciji pojavi težava med učenjem, saj hitro prihaja do izgube gradientov. Ena izmed metod, ki pospeši učenje, je naključna inicializacija Xavier, ki [5] G. E. Hinton, S. Osindero and Y.-W. Teh, "A fast uteži skalira, da se med učenjem aktivacijske funkcije ne nasičijo learning algorithm for deep belief nets," Neural tako hitro. Druge možnosti za preprečevanje nasičenih funkcij so še Batch N computation, vol. 18, pp. 1527-1554, 2006. ormalization [3] in Self-Normalizing Neural Networks [8], kjer izhode nevronov normaliziramo med učenjem. [6] A. Graves, A. r. Mohamed and G. Hinton, Predlagana inicializacija je najbolj podobna inicializacijama uteži, imenovanima Deep Autoencoder [4] ali Deep Belief Network [5], "Speech recognition with deep recurrent neural kjer je cilj, da mreži z nenadzorovanim učenjem nastavimo čim networks," in 2013 IEEE International boljše uteži, ki nam bodo pomagale optimizaciji uteži. Prednost te inicializacije je ta, da lahko faktorizacijo in učenje enonivojskih Conference on Acoustics, Speech and Signal mrež izvajamo paralelno, kar pa ni možno pri drugih dveh Processing, 2013. inicializacijah. 8. NADALJNE DELO [7] I. Sutskever, O. Vinyals and Q. V. Le, "Sequence  Paralelizacija matrične faktorizacije in učenja to Sequence Learning with Neural Networks," in enonivojskih mrež. Advances in Neural Information Processing  Izračun matrike 𝑊𝑘+1 s faktorizacijo matrike 𝑊𝑘 Systems 27, Curran Associates, Inc., 2014, pp. namesto matrike 𝑋.  3104-3112. Pretvorba inicializacije, ki bi delovala s konvolucijskimi nevronskimi mrežami. [8] G. Klambauer, T. Unterthiner, A. Mayr and S. 9. VIRI Hochreiter, "Self-Normalizing Neural Networks," CoRR, vol. abs/1706.02515, 2017. [1] A. Cutler and L. Breiman, "Archetypal analysis," Technometrics, vol. 36, pp. 338-347, 1994. [2] F. Chollet and others, Keras, GitHub, 2015. [3] S. Ioffe and C. Szegedy, "Batch normalization: Accelerating deep network training by reducing 34 Optimiranje časa in porabe goriva v modelih človeške vožnje Erik Dovgan Jaka Sodnik Fakulteta za računalništvo in informatiko Fakulteta za elektrotehniko Univerza v Ljubljani Univerza v Ljubljani Večna pot 113, 1000 Ljubljana Tržaška cesta 25, 1000 Ljubljana Odsek za inteligentne sisteme NERVteh, raziskave in razvoj, d.o.o. Institut „Jožef Stefan“ Kidričeva ulica 118, 1236 Trzin Jamova cesta 39, 1000 Ljubljana erik.dovgan@fri.uni-lj.si Ivan Bratko Bogdan Filipič Fakulteta za računalništvo in informatiko Odsek za inteligentne sisteme Univerza v Ljubljani Institut „Jožef Stefan“ Večna pot 113, 1000 Ljubljana Jamova cesta 39, 1000 Ljubljana POVZETEK Ta prispevek opisuje dvonivojski Večkriterijski optimizacij- Ko vozniki vozijo po cesti, optimirajo več kriterijev, npr. ski algoritem za iskanje strategij vožnje podobnih človeškim čas vožnje in porabo goriva. Toda teh kriterijev se nava- (MOHDS), ki vključuje modele za oponašanje človeške vo- dno ne upošteva pri gradnji modelov človeške vožnje. Za žnje ter optimira čas vožnje, porabo goriva in človeškost namen sočasne optimizacije tako človeških vidikov vožnje vožnje. Algoritem na spodnjem nivoju vključuje množico kot kriterijev vožnje smo razvili Večkriterijski optimizacij- matematičnih modelov, ki oponašajo človeško vožnjo. Algo- ski algoritem za iskanje strategij vožnje podobnih človeškim ritem na zgornjem nivoju pa je večkriterijski optimizacijski (ang. Multiobjective Optimization algorithm for discovering algoritem, razvit na podlagi algoritmov Non-dominated Sor- Human-like Driving Strategies, MOHDS). Algoritem vklju- ting Genetic Algorithm II (NSGA-II) [1] in Differential Evo- čuje modele človeške vožnje in optimira tri kriterije: čas lution for Multiobjective Optimization (DEMO) [10], ki išče vožnje, porabo goriva in podobnost s človeškimi vožnjami. najboljše vrednosti parametrov za algoritem na spodnjem MOHDS smo ovrednotili na treh cestah, ki so vključevale nivoju. Prispevek predstavi tudi okolje za simulacijo vožnje, ovinke, naklone, druga vozila in avtocesto. Dobljene strate- s katerim smo vrednotili algoritem MOHDS. gije vožnje smo primerjali s človeškimi strategijami vožnje. Rezultati kažejo, da MOHDS najde strategije vožnje, ki so z Prispevek je nadalje organiziran kot sledi. Razdelek 2 opi- vidika kriterijev primerljive s človeškimi strategijami vožnje suje sorodno delo na področju avtonomne vožnje. Okolje za v večini obravnavanih scenarijev vožnje. simulacijo vožnje je predstavljeno v razdelku 3. Razdelek 4 opisuje algoritem MOHDS. Poskusi in rezultati so navedeni Ključne besede v razdelku 5. Prispevek zaključimo s povzetkom opravlje- nega dela in napovedjo nadaljnjega dela v razdelku 6. večkriterijska optimizacija, človeške strategije vožnje, čas vo- žnje, poraba goriva 2. SORODNO DELO 1. UVOD Človeške strategije vožnje lahko posnemamo z uporabo mo- Avtonomna vožnja vozil je zelo aktivno raziskovalno podro- delov človeške vožnje, pri čemer se obstoječi modeli osredo- čje, na katerem deluje mnogo znanih podjetij, kot sta Google točajo na specifične aktivnosti vožnje, kot so sledenje vozi- [11] in Toyota [9]. Veliko sistemov za pomoč voznikom, kot lom, prosta vožnja, prehitevanje, sprememba pasu itd. Mo- je npr. sistem za ohranjanje voznega pasu, je že vgrajenih v deli za sledenje vozilom opisujejo aktivnost sledenja pred- sodobna vozila. Poleg tega popolnoma avtonomna vozila že hodnim vozilom na istem pasu. Ti modeli predpisujejo, da vozijo po javnih cestah [7]. sledeče vozilo pospešuje oziroma zavira kot odziv na podatke iz okolice, pri čemer se modeli razlikujejo v upoštevanju teh Sistemi za avtonomno vožnjo se osredotočajo na zaznavanje podatkov. Na splošno lahko podatki vključujejo hitrost in okolice, kar vključuje druga vozila, pešce, obliko ceste, razne pospešek vozila, relativno hitrost glede na predhodno vozilo, ovire na cesti itd. Toda dobljena vožnja lahko ne zadošča razdaljo do predhodnega vozila itd. [8, 14]. Modeli za prehi- ostalim kriterijem vožnje, kot so čas vožnje, poraba goriva in tevanje na regionalni cesti in spremembo pasu na avtocesti posledično onesnaževanje okolja, udobje, podobnost s člove- opisujejo odločitveni proces za ustrezne aktivnosti vožnje. ško vožnjo itd. Ti kriteriji vplivajo na sprejemljivost avtono- Spremembo pasu običajno modelirajo z modelom želje po mne vožnje s strani potnikov. Na primer, potniki ne želijo, spremembi pasu, modelom za sprejem vrzeli in modelom za da bi bila avtonomna vožnja preveč nenavadna, zelo različna izbiro vrzeli [13]. Prehitevanje modelirajo z modelom želje od njihove vožnje ali pa slabša od človeške vožnje [12]. po prehitevanju in modelom za sprejem vrzeli [3]. Ti modeli 35 so organizirani zaporedno, tj. za spremembo pasu najprej ki sočasno oponaša človeško vožnjo in optimira čas vožnje, preverijo željo po spremembi pasu, ob zadostni želji preve- porabo goriva in različnost od človeške vožnje. MOHDS se- rijo, ali je vrzel zadostna, ter če je zadostna vsaj ena vrzel, stoji iz dveh nivojev. Na spodnjem nivoju je implementira- izberejo najustreznejšo vrzel. Podoben postopek je tudi pri nih več modelov človeške vožnje, ki vodijo vozilo v različnih prehitevanju, le da ne vključuje izbire vrzeli. aktivnostih vožnje. Vrednosti parametrov modelov na spo- dnjem nivoju iščemo z algoritmom na zgornjem nivoju, tj. Modeli človeške vožnje posnemajo človeško obnašanje, pri večkriterijskim optimizacijskim algoritmom, ki minimizira čemer pa zanemarijo ostale kriterije, ki so tudi pomembni čas vožnje, porabo goriva in različnost od človeških strategij med vožnjo, kot so čas vožnje in poraba goriva. Razvitih je vožnje. Začetna verzija algoritma je bila predstavljena v [2]. bilo več pristopov za optimiranje teh kriterijev, ki večinoma Ta verzija je sedaj nadgrajena z modeli za vožnjo po klancih vključijo vse kriterije v eno kriterijsko funkcijo ali pa opti- in ovinkih, modeli za spreminjanje pasu ter s primerjavo s mirajo samo porabo goriva, čas vožnje pa vključijo kot ome- človeškimi strategijami vožnje oziroma optimizacijo glede na jitev. Hellstrom in sod. [4] so razvili metodo dinamičnega različnost od človeških strategij vožnje. programiranja, ki optimira uteženo vsoto kriterijev. Razvi- tih je bilo tudi več analitičnih metod za optimiranje utežene 4.1 Algoritem na spodnjem nivoju vsote kriterijev [6] oziroma le porabe goriva [5]. Algoritem na spodnjem nivoju vključuje množico matema- tičnih modelov, ki oponašajo človeško vožnjo in upravljajo Obstoječe metode za iskanje strategij vožnje se osredotočajo vozilo pri naslednjih aktivnostih vožnje: (a) prosta vožnja, bodisi na strategije vožnje podobne človeškim, bodisi na op- (b) sledenje vozilom, (c) zaviranje v sili, (d) prehitevanje in timizacijo časa vožnje, porabe goriva in/ali ostalih kriterijev. (e) sprememba pasu. Modeli za sledenje vozilom, prosto vo- Ta prispevek predstavlja algoritem, ki rešuje oba problema žnjo in zaviranje v sili določajo pospešek vozila, medtem ko hkrati: modelira človeško vožnjo z modeli človeške vožnje modeli za prehitevanje in spremembo pasu odločajo, kdaj ter uglašuje parametre teh modelov, pri čemer optimira čas vozilo spremeni vozni pas. vožnje, porabo goriva in podobnost s človeškimi vožnjami. Model za sledenje vozilom temelji na modelu Gazis-Herman- 3. SIMULACIJA VOŽNJE Rothery (GHR) [8], ki določa pospešek vozila glede na hi- Simulacijsko okolje je namenjeno vrednotenju strategij vo- trost vozila, razdaljo do predhodnega vozila in razliko hitro- žnje in omogoča simulacijo vožnje po regionalni cesti in dvo- sti med vozilom in predhodnim vozilom. Ko je vozilo daleč pasovni avtocesti. Cesta je razdeljena na odseke, pri čemer od predhodnega vozila, uporabimo namesto modela za sle- za vsak odsek določimo dolžino, omejitev hitrosti, polmer denje model za prosto vožnjo, ki vodi vozilo s konstantnim ovinka, naklon, smer vožnje pasov in možnost prehitevanja. pospeškom, dokler ni dosežena ciljna hitrost. Ko je vozilo Na cesti so lahko druga vozila, ki vozijo v smeri pasu ter ne preblizu predhodnemu vozilu, uporabimo model za zavira- prehitevajo. Tem vozilom določimo hitrost in razdaljo do nje v sili. Poleg zgornjih modelov uporabimo za določanje predhodnih vozil. pospeška še naslednje omejitve. Pospešek se lahko zmanjša glede na omejitve vozila, kot je opisano v razdelku 3. Za Vozilo vodimo z ukrepi vodenja, ki vključujejo pospešek, kot vsak vozni pas ima vozilo določeno ciljno hitrost. Poleg tega vozila glede na smer ceste in prestavo. Prestava se spreminja se ciljna hitrost voznega pasu zmanjša, če je cestni odsek glede na spodnjo in zgornjo mejo hitrosti motorja. Pri vožnji ovinek ali klanec, pri čemer je stopnja zmanjšanja ciljne hi- vozila upoštevamo tudi fizikalne omejitve vozila, s katerimi trosti odvisna od ostrine ovinka in naklona odseka. določimo največji pospešek, ki ga lahko vozilo doseže. Prehitevanje na regionalni cesti je določeno z modeloma že- Simulacijo izvajamo v več korakih, dokler vožnja po celotni lje po prehitevanju in za sprejem vrzeli [3]. Model želje po cesti ni končana. V vsakem koraku preverimo veljavnost vo- spremembi pasu na podlagi ciljne hitrosti, razdalje do pred- žnje, tj. vozilo se ne sme ustaviti, se ne sme zaleteti in ne hodnega vozila in hitrosti predhodnega vozila določi, kdaj sme kršiti omejitve hitrosti. Za celotno vožnjo izračunamo vozilo želi prehiteti. Nato preverimo sprejemljivost prednje naslednje kriterije: čas vožnje, porabo goriva in podobnost s vrzeli na drugem pasu na podlagi hitrosti vozila, hitrosti človeško vožnjo. Porabo goriva izračunamo na podlagi dia- predhodnega vozila na istem pasu in hitrosti predhodnega grama specifične porabe goriva. Podobnost s človeško vožnjo vozila na drugem pasu. izračunamo na podlagi vnaprej pridobljenih podatkov o vo- žnjah voznikov. Ker pa želimo minimizirati vse kriterije, Sprememba pasu na avtocesti je določena z modelom želje po namesto podobnosti izračunamo različnost glede na člove- spremembi pasu in modelom za sprejem vrzeli [13]. Podobno ško vožnjo. Za vsakega voznika izračunamo različnost s sre- kot pri prehitevanju tudi ta model želje po spremembi pasu dnjim kvadratnim odklonom (ang. Root Mean Square Error, na podlagi hitrosti vozila, razlike hitrosti glede na predho- RMSE) in vrnemo RMSE voznika z najmanj različno vožnjo. dno vozilo in razlike hitrosti glede na predhodno vozilo na Za ta izračun upoštevamo dva atributa: hitrost in odmik od drugem pasu določi, kdaj vozilo želi spremeniti vozni pas. sredine desnega voznega pasu. Pri tem izračunamo RMSE Nato preverimo sprejemljivost prednje in zadnje vrzeli na vsakega atributa in upoštevamo povprečje. drugem pasu glede na razlike hitrosti glede na predhodno vozilo in vozilo zadaj na drugem voznem pasu. 4. ALGORITEM ZA ISKANJE STRATEGIJ 4.2 Algoritem na zgornjem nivoju VOŽNJE PODOBNIH ČLOVEŠKIM Algoritem na zgornjem nivoju je večkriterijski optimizacij- Ta razdelek opisuje Večkriterijski optimizacijski algoritem ski algoritem, razvit na podlagi algoritmov NSGA-II [1] in za iskanje strategij vožnje podobnih človeškim (MOHDS), DEMO [10]. Algoritem išče najboljše vrednosti parametrov 36 0.6 0.55 0.6 0.5 0.55 0.45 0.5 0.4 0.45 0.35 0.4 0.35 7 Različnost od človeške vožnje [RMSE] Različnost od človeške vožnje [RMSE] 7 6 3 5 4 1300 1500 1700 1900 Poraba goriva [l] 1300 1500 1700 1900 3 Poraba goriva [l] Čas vožnje [s] Čas vožnje [s] Slika 1: Strategije vožnje v kriterijskem prostoru, najdene z algoritmom MOHDS: dva različna pogleda. modelov na spodnjem nivoju, pri čemer minimizira čas vo- sti smo ocenjevali na prvi cesti in prvi polovici druge ceste, žnje, porabo goriva in različnost od človeških voženj. Za is- medtem ko smo različnost po odmiku od sredine desnega kanje uporablja evolucijski pristop, ki množico rešitev izbolj- voznega pasu preverjali na drugi polovici druge ceste ter na šuje skozi več generacij. Za ohranjanje konstantne velikosti tretji cesti. Vsako strategijo vožnje smo vrednotili na vseh populacije uporablja nedominirano razvrščanje in metriko treh cestah in pri iskanju strategij optimirali skupni čas vo- nakopičenosti iz algoritma NSGA-II [1]. Rezultat optimiza- žnje, skupno porabo goriva in skupno različnost od človeške cije je množica nedominiranih rešitev. Za več podrobnosti vožnje. Večkriterijsko optimizacijo smo izvajali skozi 200 ge- glej [2]. neracij, pri čemer je bila velikost populacije 100 rešitev. Za ostale parametre algoritma MOHDS glej [2]. 5. POSKUSI IN REZULTATI MOHDS smo vrednotili na treh cestah in rezultate primerjali 5.2 Rezultati s človeškimi vožnjami. Naslednji razdelki opisujejo testne Strategije vožnje, pridobljene z algoritmom MOHDS, so pri- scenarije in dobljene rezultate. kazane na sliki 1. Rezultati kažejo, da MOHDS najde stra- tegije vožnje z različnimi kompromisi med kriteriji, tj. časom 5.1 Opis poskusov vožnje, porabo goriva in različnostjo od človeške vožnje. Po- Vrednotenje algoritma MOHDS smo izvedli na naslednjih leg tega se strategije vožnje, ki imajo bodisi najkraši čas cestah: (a) Regionalna cesta dolžine 11.450 m, na kateri ni vožnje bodisi najnižjo porabo goriva, najbolj razlikujejo od drugih vozil. Cesta vključuje različne omejitve hitrosti, tri človeške vožnje. stopnje ovinkov (od blagega do zelo ostrega) ter tri stopnje klancev navzgor in navzdol (od blagega do zelo strmega); Slika 2 prikazuje primerjavo človeških strategij in strategij (b) Dvopasovna regionalna cesta dolžine 9.000 m, po kateri algoritma MOHDS glede na čas vožnje in porabo goriva. vozijo tudi druga vozila. Cesta je brez ovinkov in klancev Kljub temu, da smo optimirali skupni čas vožnje, skupno ter ima konstantno omejitev hitrosti. Vozila na desnem pasu porabo goriva in skupno različnost od človeške vožnje za vse spreminjajo hitrost, medtem ko imajo vozila na levem pasu, tri ceste, so na tej sliki prikazane delne vrednosti kriterijev za ki vozijo v nasprotno smer, konstantno hitrost. Razdalja vsako cesto posebej. Posledično so lahko določene strategije med vozili na levem pasu je na začetku ceste krajša, nato slabe na posamezni cesti, a so vseeno nedominirane glede pa daljša. Cesta se začne s polno sredinsko črto, na pri- na skupne vrednosti kriterijev za vse tri ceste. Ti rezultati bližno polovici ceste pa sredinska črta postane prekinjena; kažejo, da MOHDS najde strategije vožnje, ki so primerljive (c) Dvopasovna avtocesta dolžine 10.000 m, po kateri vozijo s človeškimi strategijami vožnje na prvi in tretji cesti. Poleg tudi druga vozila. Cesta je brez ovinkov in klancev ter ima tega kažejo, da MOHDS uspešneje optimira porabo goriva, konstantno omejitev hitrosti. Na celotni cesti je prekinjena tj. najde strategije vožnje z nižjo porabo goriva v primerjavi sredinska črta. Vozila na desnem pasu spreminjajo hitrost, z vozniki. Dodatna analiza strategij vožnje algoritma MO- medtem ko imajo vozila na levem pasu, ki vozijo v isto smer, HDS na drugi cesti je pokazala, da strategije prehitevajo konstantno hitrost. Razdalja med vozili na levem pasu je na bolj redko kot ljudje, kar onemogoča dodatno krajšanje časa začetku ceste krajša, nato pa daljša. vožnje. Za vsako cesto smo pridobili podatke o človeških vožnjah 6. ZAKLJU ČEK 30 voznikov, pri čemer je vsak voznik prevozil vsako cesto V prispevku smo predstavili dvonivojski Večkriterijski op- dvakrat. Različnost od človeške vožnje smo nato izračunali timizacijski algoritem za iskanje strategij vožnje podobnih za vsakega voznika posebej kot povprečje vseh voznikovih človeškim (MOHDS), ki sočasno optimira čas vožnje, po- voženj, kot je opisano v razdelku 3. Različnost po hitro- rabo goriva in človeškost vožnje. Algoritem smo vrednotili z 37 MOHDS 8. LITERATURA vozniki (a) [1] K. Deb, A. Pratap, S. Agrawal in T. Meyarivan. A 2.4 fast and elitist multiobjective genetic algorithm: 2.2 NSGA-II. IEEE Transactions on Evolutionary 2 Computation, 6(2):182–197, 2002. 1.8 [2] E. Dovgan, J. Sodnik, I. Bratko in B. Filipič. Multiobjective discovery of human-like driving 1.6 strategies. V GECCO’17 Companion – Proceedings of 1.4 the Genetic and Evolutionary Computation Poraba goriva [l] 1.2 Conference GECCO 2017, 8 strani, 2017. 1 [3] H. Farah in T. Toledo. Passing behavior on two-lane 450 500 550 600 650 700 750 800 850 900 highways. Transportation Research Part F, Čas vožnje [s] 13(6):355–364, 2010. (b) 2.8 [4] E. Hellstrom, J. Aslund in L. Nielsen. Design of an 2.6 efficient algorithm for fuel-optimal look-ahead control. 2.4 Control Engineering Practice, 18(11):1318–1327, 2010. 2.2 2 [5] P. G. Howlett, P. J. Pudney in X. Vu. Local energy 1.8 minimization in optimal train control. Automatica, 1.6 1.4 45(11):2692–2698, 2009. 1.2 [6] M. Ivarsson, J. Aslund in L. Nielsen. Optimal speed 1 Poraba goriva [l] 0.8 on small gradients – consequences of a non-linear fuel 0.6 map. V Proceedings of the 17th World Congress of the 420 440 460 480 500 520 540 560 580 600 620 International Federation of Automatic Control Čas vožnje [s] IFAC’08, strani 3368–3373, 2008. (c) 2.4 [7] E. Jaffe. The First Look at How Google’s Self-Driving 2.2 Car Handles City Streets, 2014. Dostopno na: 2 https://www.citylab.com/life/2014/04/first- 1.8 look-how-googles-self-driving-car-handles- 1.6 city-streets/8977/. 1.4 [8] Y. Li in D. Sun. Microscopic car-following model for 1.2 the traffic flow: the state of the art. Journal of 1 Control Theory and Applications, 10(2):133–143, 2012. Poraba goriva [l] 0.8 [9] R. Read. Toyota will roll out autonomous cars by the 260 280 300 320 340 360 380 400 420 440 460 480 ’mid-2010s’, 2013. Dostopno na: Čas vožnje [s] http://www.thecarconnection.com/news/1087636\ _toyota-will-roll-out-autonomous-cars-by-the- mid-2010s. Slika 2: Primerjava strategij vožnje algoritma MO- HDS in voznikov glede na čas vožnje in porabo go- [10] T. Robič in B. Filipič. DEMO: Differential evolution riva na (a) prvi, (b) drugi in (c) tretji cesti. for multiobjective optimization. V Proceedings of the Third International Conference on Evolutionary Multi-Criterion Optimization – EMO 2005, zvezek 3410, strani 520–533, 2005. vožnjo po treh cestah, tj. prazni regionalni cesti, regionalni [11] R. J. Rosen. Google’s self-driving cars: 300,000 miles cesti z drugimi vozili in avtocesti z drugimi vozili. Dobljene logged, not a single accident under computer control, strategije vožnje smo primerjali z vožnjami voznikov. Rezul- 2012. Dostopno na: tati kažejo, da MOHDS najde strategije vožnje z različnimi http://www.theatlantic.com/technology/archive/ kompromisi med kriteriji. Poleg tega MOHDS najde po- 2012/08/googles-self-driving-cars-300-000- dobne strategije vožnje kot vozniki na prazni regionalni ce- miles-logged-not-a-single-accident-under- sti in na avtocesti, medtem ko na regionalni cesti z ostalimi computer-control/260926/. vozili najde počasnejše strategije vožnje. V bodoče bomo [12] B. Schoettle in M. Sivak. A Survey of Public Opinion dodatno analizirali in izboljševali algoritem MOHDS pred- about Autonomous and Self-Driving Vehicles in the vsem za vodenje vozil po regionalni cesti, kjer vozijo tudi U.S., the U.K., and Australia. Technical Report druga vozila in je prehitevanje dovoljeno. UMTRI-2014-21, Transportation Research Institute, The University of Michigan, 2014. 7. ZAHVALA [13] T. Toledo, H. N. Koutsopoulos in M. E. Ben-Akiva. Projekt ( Večkriterijsko iskanje strategij vožnje za avtono- Modeling integrated lane-changing behavior. ” mna vozila“, št. Z2-7581) je sofinancirala Javna agencija za Transportation Research Record, 1857:30–38, 2003. raziskovalno dejavnost Republike Slovenije iz državnega pro- [14] T. H. Yang in C. W. Zu. Linear dynamic car-following računa. Delo predstavljeno v tem prispevku je delno finan- model. V Proceedings of the 5th World Congress on ciralo tudi podjetje NERVteh, raziskave in razvoj, d.o.o. Intelligent Control and Automation, zvezek 6, strani 5212–5216, 2004. 38 User-friendly Multi-objective Learning of Accurate and Comprehensible Hybrid-trees Benjamin Novak Rok Piltaver Matjaž Gams Ježef Stefan Institute, Ježef Stefan Institute, Ježef Stefan Institute, Department of Intelligent Department of Intelligent Department of Intelligent Systems Systems Systems Jamova cesta 39 Jamova cesta 39 Jamova cesta 39 1000, Ljubljana 1000, Ljubljana 1000, Ljubljana benjaminovak@gmail.com rok.piltaver@ijs.si matjaz.gams@ijs.si ABSTRACT models such as support vector machines, artificial neural In data-mining applications, users must often choose be- networks and ensembles (e.g. random forest, boosting and tween a comprehensible classifiers with low accuracy or a stacking algorithms) that achieve higher accuracy then clas- more accurate but incomprehensible classifier. A possible sification trees in many domains but are not comprehensi- solution is to apply MOLHC algorithm, which finds the com- ble, hence they are referred to as black-box classifiers. Users plete set of non-dominated hybrid-trees according to accu- are therefore faced with a difficult decision: they can either racy and comprehensibility. However, tools that would help choose a comprehensible classification tree with relatively the user select an appropriate hybrid-tree from the set are low accuracy or an incomprehensible classifier with relatively missing. Therefore, we present MOLHC implemented as high accuracy. In many cases none of the two options is suf- an add-on for Orange data-mining suite. We implemented ficient. widgets for learning, evaluation and visualization of individ- ual hybrid-trees and the set of non-dominated hybrid-trees. The difficulty of learning accurate and comprehensible clas- They comprise a user-friendly toolbox that enables users to sifiers arises from the fact that the two objectives must be efficiently execute the multi-objective data-mining process. considered as equally important and that they are conflict- ing: increasing one often decreases the other. The solution 1. INTRODUCTION is multi-objective learning. It is based on multi-objective There are two key criteria for selecting the classifier that optimization and therefore returns a set of non-dominated will be deployed in a data-mining application: predictive classifiers (not a single classifier). This enables the user performance and comprehensibility. The most appropriate to improve the decision on the accuracy-comprehensibility measures of predictive performance (e.g. accuracy, area un- trade-off by deferring it: instead of taking an arbitrary deci- der the ROC curve, sensitivity, specificity) is used to select a sion about the relative importance of the learning objective subset of acceptable classifiers depending on the objectives before the execution of the learning algorithm (and having of the application. However, acceptable predictive perfor- to rerun it with different settings to obtain a classifier with mance is often not enough. Comprehensibility is reported a different trade-off) user can now take a well informed de- as decisive factor for classifier application in domains such cision by simply comparing the subset of all non-dominated as: medicine, credit scoring, churn prediction, and bioinfor- classifiers returned by the multi-objective algorithm. matics [2]. It is described as ”the ability to understand the logic behind a prediction of the model” [4] or as ”the ability This paper focuses on a recently developed example of such to understand the output of induction algorithm” [3]. It is algorithm named Multi-objective learning of hybrid classi- important because comprehensible classifiers enable expla- fiers (MOLHC) [7]. It finds the entire Pareto set of hybrid- nation for classification of each instance, classifier valida- trees (according to accuracy and comprehensibility) by re- tion, knowledge discovery and support classifier generaliza- placing sub-trees in the initial classification tree (given as tion improvement. Classification trees for example are often an input) with black-box classifiers (also given as an input). used in machine-learning applications because they are one The resulting hybrid-trees consist of regular leaves contain- of the most comprehensible classification models. ing easily comprehensible rules and incomprehensible black- box leaves enabling improved accuracy. MOLHC was shown On the other hand, there are more complex classification to produce classifiers with accuracy-comprehensibility trade- offs and offer insights into the data-mining task that are not available using traditional machine-learning algorithms as well as being fast and simple enough for applications in real-world problems [5]. After finding the entire Pareto front, the user has to se- lect the best hybrid-tree, which was a cumbersome and time consuming task due to the lack of user-friendly tools to sup- port it. Therefore, this paper presents a user-friendly im- 39 plementation of the MOLHC algorithm and additional tools with a graphical user interface (GUI), which make learning, comparing and selecting the hybrid-trees and evaluating the results of learning efficient. The algorithm and the corre- sponding tools are implemented as an Orange add-on with three new Orange widgets shown in the Figure 1. The wid- gets are components in the visual programming environment called Orange Canvas [1]. Each widget offers a self contained data-mining functionality and a graphical user interface for setting its parameters and presenting its (main) results. A widget can be connected to other widgets by passing data from its output(s) to the input communication channel(s) of the other widgets, which enables visual programming of data-mining applications. Figure 2: MOLCH widget shows the Pareto front with 6 hybrid-trees (the chosen hybrid-tree is marked with an orange circle). • Classification tree: the initial tree used by MOLCH. • Black-box classifier: an (incomprehensible) classifier that has a higher accurate than the initial tree. Figure 1: The list of developed Orange widgets. The following sections describe the MOLHC related Orange widgets in the order in which they are typically applied dur- ing the data-mining process. Section 2 presents the MOLHC widget, which implements the MOLHC algorithm and visu- alizes the Pareto front of learned hybrid-trees. Section 3 presents the widget that visualizes the hybrid-tree chosen on the Pareto set and Section 4 presents the widget used for the evaluation of MOLHC results. Section 5 concludes the paper and discusses the ideas for future work. 2. MOLHC WIDGET Figure 3: Connecting the MOLHC widget with the The MOLHC widget implements Multi-objective learning of other widgets providing the required inputs (on the classifiers (MOLHC) algorithm. The widget finds and visu- left) and accepting its output (on the right). alizes the entire Pareto-optimal set of hybrid-trees accord- ing to accuracy and comprehensibility as shown in Figure 2. Comprehensibility is defined as the share of instances that The output of the MOLHC widget is the hybrid-tree chosen fall into the regular (not black-box) leaves of the hybrid-tree by clicking on it in the Pareto-front viewer. The MOLHC - this measure was designed based on the results of survey widget offers the following options: on the comprehensibility of classification trees [8]. • Splitting options: splitting of the input data set into Pareto-front viewer, which is a part of the widget’s GUI (the the training and testing set (used to estimate the accu- right part in Figure 2), supports comparison and selection racy and comprehensibility of the hybrid-trees on the of an appropriate hybrid-tree from the set of learned hybrid- Pareto-front) according to the set proportion or using trees and enables extracting insights about the data-mining all the data for both testing and training. domain as discussed in our previous work [6]. • Use local black-box classifier: use multiple black-box The MOLHC widget is compatible with the standard Orange classifiers (each trained on the instances belonging to a widgets so they can be used together as explained below and specific leaf in the initial tree) instead of a single black- illustrated in Figure 3. Its inputs are: box classifier trained on all the training instances. • Visualization options: zoom, selection and pan tool, • Data: data set for training and testing. change size or opacity of hybrid-tree symbols. 40 inputs of the MOLHC widget, except that multiple black- box classifiers can be provided for comparison. The parameters of MOLHC evaluation widget are: • The number of folds used for cross-validation. • Train data proportion: the percentage of instances from the data set to be used as the training set, the rest of the instances are used as the testing set. • Use local black-box classifier: use multiple black-box classifiers (one per leaf in the initial tree) instead of a single black-box classifier trained on all the training instances. Figure 6 show an example of MOLHC evaluation. It is based on several measures for each provided black-box classifiers: Figure 4: Evaluation widget with inputs. • Method: the name of the black-box classifier. 3. HYBRID-TREE VIEWER • BB accuracy: the accuracy of the black-box classifier. The hybrid-tree viewer widget visualizes a hybrid-tree (an • Tree accuracy: the accuracy of the initial tree. example is shown in Figure 5) received as an input from the MOLHC widget. The visualization of a hybrid-tree is used • Baseline hypervolume: the hypervolume for the set of to validate it or extract knowledge from it. In addition, a two baseline solutions (the initial tree and black-box pair of hybrid-tree viewer widgets positioned side by side classifier). can be used to compare a pair of hybrid-trees according to their structure; comparing them according to the accuracy • MOLHC hypervolume: the hypervolume for the Pareto and comprehensibility is done with the Pareto-front viewer set of hybrid-trees learned by the MOLHC algorithm. in the MOLHC widget. • Hypervolume difference: the difference between the MOLHC and the baseline hypervolume. The parameters of the Hybrid-tree viewer define the display options, which are similar as the ones available in the stan- dard Orange tree viewer. In addition, the user can choose The results provided by the MOLHC evaluation widget are whether to use the train or the test data set to compute the interpreted as follows. If the difference between the best statistical data shown in the tree nodes. BB accuracy and the tree accuracy is small, the user should consider using a regular classification tree instead of a black- The hybrid-tree viewer widget provides two outputs that box classifier or a hybrid-tree. In general a higher BB ac- depends on which node of the hybrid-tree the user selected curacy means that the corresponding black-box classifier by clicking on it: has a higher potential to increase the accuracy of the ini- tial classification tree. However, the actual results of the • MOLHC algorithm depends on the accuracy of the black- Selected data: the subset of the data set instances that box classifier in each leaf of the initial tree, therefore the belong to the selected tree node. best black-box classifier should be selected according to the • Selected black-box model: the black-box classifier if a MOLHC hypervolume. The overall success of the MOLHC black-box leaf is chosen. algorithm is measured by the hypervolume difference - the higher the better. Finally, the user should select an ap- 4. MOLHC EVALUATION propriate hybrid-tree regardless of the hypervolume differ- The MOLHC evaluation widget compares the Pareto set of ence because it shows only the advantage of the MOLHC hybrid-trees (i.e. the output of the MOLHC widget) with approach over the baseline algorithms, which depend on all the set of baseline solutions consisting of the initial classifi- the hybrid-trees in the Pareto set and is therefore not appro- cation tree and the black-box classifier. It is used to evaluate priate for the evaluation of a single hybrid-tree. To select the results of multi-objective learning and to select the best a single hybrid-tree, the user should use Pareto front and algorithm for learning the black-box classifier (used as the hybrid-tree visualizer instead. input to the MOLHC widget/algorithm). The evaluation is based on the hyper-volume measure [9], which is often 5. CONCLUSION used to compare results of multi-objective optimization al- The paper presents an implementation of the MOLHC algo- gorithms. rithm, which is a multi-objective learning algorithm that finds the complete set of non-dominated hybrid-trees ac- The MOLHC evaluation widget requires at least three inputs cording to accuracy and comprehensibility. The algorithm (an example is shown in Figure 4). They are the same as te is most suitable for the data-mining applications where a 41 Figure 5: Hybrid-tree viewer widget: the black-box leaves are marked with black. Figure 6: MOLHC evaluation widget with calculated results. regular classification tree is not accurate enough and black- [3] R. Kohavi. Scaling up the accuracy of naive-bayes box classifier (with a higher accuracy) is not acceptable be- classifiers: A decision-tree hybrid. In KDD, volume 96, cause it is not comprehensible. We implemented MOLHC as pages 202–207, 1996. an add-on for Orange, which is a popular and user-friendly [4] D. Martens, J. Vanthienen, W. Verbeke, and data-mining suite that offers visual programming and rich B. Baesens. Performance of classification models from a visualizations. We implemented three new Orange widgets user perspective. Decision Support Systems, for learning and visualizing the set of non-dominated hybrid- 51(4):782–793, 2011. trees, visualizing individual hybrid-trees and for evaluation [5] R. Piltaver. Constructing Comprehensible and Accurate of the MOLHC results. They comprise a user-friendly tool- Classifiers Using Data Mining Algorithms. PhD thesis, box that enables users to efficiently execute the multi-ob Jožef Stefan international postgraduate school, 8 2016. jective data-mining process. Nevertheless, we are consid- [6] R. Piltaver, M. Luštrek, and M. Gams. Multi-objective ering several improvement that could make the developed learning of accurate and comprehensible classifiers – a widgets more user friendly. We plan to improve the doc- case study. In Proceedings of 7th European Starting AI umentation in order to make our work more accessibly to Researcher Symposium â ˘ A¸ S STAIRS 2014, pages a wider user base and reduce the learning curve. Another 220–229). IOS Press, 2014. improvement would be to color similar hybrid-trees on the [7] R. Piltaver, M. Luštrek, J. Zupančič, S. Džeroski, and Pareto front according to Hammilton distance which would M. Gams. Multi-objective learning of hybrid classifiers. help the user when comparing them. Finally, an improve- In Proceedings of the Twenty-first European Conference ment of the Hybrid-tree visualizer widget would add an op- on Artificial Intelligence, pages 717–722. IOS Press, tion to replace the sub-trees that have only black-box leaves 2014. with a single black-box leaf. [8] R. Piltaver, M. Luštrek, M. Gams, and S. Martinčić-Ipšić. What makes classification trees 6. REFERENCES comprehensible? Expert Systems with Applications, [1] J. Demšar, B. Zupan, G. Leban, and T. Curk. Orange: 62(C):333–346, Nov. 2016. From experimental machine learning to interactive data [9] E. Zitzler, L. Thiele, M. Laumanns, C. M. Fonseca, and mining. Knowledge discovery in databases: PKDD V. G. Da Fonseca. Performance assessment of 2004, pages 537–539, 2004. multiobjective optimizers: An analysis and review. [2] A. A. Freitas. Comprehensible classification models: a IEEE Transactions on evolutionary computation, position paper. ACM SIGKDD explorations newsletter, 7(2):117–132, 2003. 15(1):1–10, 2014. 42 Automatic Tennis Analysis with the use of Machine Learning and Multi-Objective Optimization Miha Mlakar Mitja Luštrek Jožef Stefan Institute Jožef Stefan Institute Jamova cesta 39 Jamova cesta 39 1000 Ljubljana, Slovenia 1000 Ljubljana, Slovenia mitja.lustrek@ijs.si miha.mlakar@ijs.si ABSTRACT the effort, distance covered, sprints analysis and much more. Here, Wearable devices for monitoring players’ movements are heavily the leading provider in the world is Catapult Sports, whose S5 used in many sports. However, the existing commercial and product is currently used by the best tennis player in the world research sports wearables are either not tennis-specific, or are Andy Murray. Nevertheless, the problem with S5 is that it offers worn on the wrist or in the racquet and thus offer too limited no tennis-specific metrics. That is why in our research we add information. We therefore added tennis-specific information to a tennis-specific information to the metrics already available in the leading commercial device. Our solution is two-fold. Firstly, we Catapult S5 system, to produce a comprehensive solution that developed a model for classifying shot types into forehand, enables professional players to make better tactical preparations backhand and serve. Secondly, we designed an algorithm based and to improve their game. on multi-objective optimization to distinguish active play from the Our algorithm consists of two parts. In the first part, we detect time in-between points. By combining both parts with the general when a tennis shot occurs and which type of shot it is. In the movement information already provided by the device, we get a second part, we focus on detecting when the players actually play comprehensive set of metrics that are used by professional tennis players and coaches to objectively measure a player’s points (active play) and when they are in-between points. This allows us to determine the actual net playing time and real performance and enable in-depth tactical analysis. distance covered and also adds context to shots which enable complex analysis like “Is the player playing weaker shots, if the Categories and Subject Descriptors point is longer than 15 seconds?” With this solution the players I.2.6. Artificial Intelligence: Training and their coaches get a continuous comprehensive view of the player’s game, both the physical and the tactical part of it. General Terms Algorithms, Measurement, Experimentation 2. DATA AQUISITION To obtain sensor data we used the commercially available S5 Keywords device from Catapult. The position of the device was high on the player’s back attached to a tight shirt. The device contains a 3D Tennis; Wearable analytics; Shot detection; Optimization; accelerometer (frequency 100 Hz), 3D gyroscope (frequency 100 Hz), 3D magnetometer (frequency 100 Hz) and GPS sensor, 1. INTRODUCTION returning latitude and longitude (frequency 10 Hz). The use of wearable sensors in sport is growing fast and can We recorded 5 different professional tennis players for 6 hours in already be considered essential for success in some disciplines. In total. Due to the 100 Hz frequency, we obtained 2,172,363 data tennis the analytics started with computer vision and sensors for records. In this time, we recorded 1,373 shots. Each shot was measuring shots. However, both of these approaches have labeled as a serve, forehand or backhand. As for detecting active limitations for professional use. The sensors worn on the playing play, we also manually labeled the beginning and end of each wrist or built into the tennis racquets deliver information about sequence of active play. Because we were interested in creating an the shots [6] or enable the analysis and modeling of different shot algorithm for detecting shot types and active plays in actual techniques [7]. However, the problem with this information is the matches, all the data were recorded during matches and none lack of context (under what circumstances and where on the court during predefined situations of practice sessions. did a specific shot occur), so it is not sufficiently actionable, i.e. cannot be used for tactical preparations or to significantly improve 3. SHOT DETECTION players' game. Video analysis offers better information, and there has been a lot of research on this topic [1, 5]. However, cheap For every data point obtained from our device, we extracted a solutions offer low accuracy, while better solutions are extremely number of features used by the shot detection algorithm. We used expensive because they require advanced cameras with complex supervised machine learning to train a model to detect shots. With software for calibration. Additionally, they are bound to a specific this model we classified every data point and evaluated the shot court, so the information is not available whenever needed by the detection. player or coach. 3.1 Feature Extraction Due to these limitations, devices worn on the torso, and equipped To define informative features for shot detection, we visualized with accelerometers, gyroscopes and GPS receivers are emerging and examined the traces for the accelerometer and gyroscope. as the new approach. These devices are perfect for determining Since we saw that every shot is associated with body rotation, our 43 main source for feature extraction was the gyroscope - more (LOPO), where we used one player’s data for testing and the data specifically angular speeds on axes 1 (Roll) and 3 (Yaw) - and not from the other players for training. This approach enables us to the accelerometer. Figure 1 shows a typical trace of the gyroscope estimate the accuracy of the models for previously unseen players and accelerometer for a backhand shot. with different shot techniques. When evaluating the models, we classified each data entry (10 ms) as a shot or no-shot, and the type of shot. With this approach almost all the data points were classified as no-shots, so calculating the classification accuracy would be useless. We therefore focused on the precision and recall. 3.3 Results The results for detecting shots and shot types for the cross- validation and for the LOPO approach are presented in Tables 1 and 2. Cross-validation LOPO Precision 97.3% 97.3% Recall 34% 96.6 % 96.5% 35% Table 1: Precision and recall for detecting tennis shots Figure 1: Gyroscope and accelerometer traces for backhand shot marked with the vertical line. Cross-validation As our main feature, we calculated a feature called Peak_strength Foreh. Backh. Serve All as follows: Precision 95.3% 94.3% 99.1% 96.2% - Calculate absolute sum of angular speeds on axes 1 and Recall 91.4% 90.2% 99.3% 93.6% 3 LOPO - Raise it to the power 4, to emphasize higher values Foreh. Backh. Serve All - Apply Butterworth band-pass filter with high and low Precision 91.5% 93.6% 99.8% 95.0% cutoff frequencies of 1.5 and 25 Hz. Recall 90.5% 90.6% 98.2% 93.1% - To get the final Peak_strength value, set the lower peak to zero when two peaks are too close (the distance was Table 2: Precision and recall for detecting types of tennis shots set by a domain expert to 1.3 s) As we can see, the precision and recall obtained with cross- High Peak_strength values calculated in this way mark potential validation and LOPO are very similar. This means that the built shots. Additionally, we calculated several other features. We set models are relatively independent from the type of player or his two different window sizes (0.8 s and 1.2 s) and calculated the technique or style of play. average values, variances and standard deviations for each The main sources of errors are fast unnatural body rotation accelerometer and gyroscope axis. We added the sums of and movements and special events that occur during the play. An differences between all pairs of gyroscope axis values and also example from our data set is a player warming up doing very between accelerometer axis values. We also calculated the speed similar body movements as during shots, or a player throwing his of movement from the GPS coordinates. To illustrate its racquet at the fence with the same body movement as when importance, Figure 2 shows how the combinations of serving. Peak_strength and speed of movement separates shots and shot attempts (high Peak_strength values that are not shots). 4. DETECTING ACTIVE PLAY The algorithm for detecting active play during a tennis game 3.2 Experimental Setup could only be developed after we have detected the shots. The We divided the evaluation in two parts. Firstly, we evaluated how reason is that we want our algorithm not only to have a high well we can detect if a shot has occurred, and secondly, we tried classification accuracy, but also to include as many shots as to detect which type of shot was made. possible in the detected active play. In other words, misdetection For building the models, after empirical comparison of several of active play is less undesirable when no shots are made. So due algorithms, we chose the Random Forest (RF) [2] algorithm. Each to having two objectives, the detection was formulated as a multi- RF consisted of 10 decision trees, the minimum number of objective optimization problem. samples required to split an internal node was 8, and the minimum number of samples required at a leaf node was 4. 4.1 Feature Extraction The main idea when detecting the starting (end ending) point of a We evaluated the models in two ways. Firstly, we performed 10- sequence of active play (rally) was that at this point the difference fold cross validation using Stratified shuffle split [3]. This between the activity before and after would be the largest. procedure ensured equal class distributions between training and test sets. Secondly, we used the leave-one-player-out approach 44 We used accelerometer values because they better represent the finished. To include even such shots in the detected rallies, we players’ movement then gyroscope values, which primarily spike would need to sacrifice a lot of classification accuracy. when making shots. From these values, we calculated a modified variance that gives more emphasis to the largest variations in data traces: So for each data point, we calculated three additional features based on var*: the back overall variance (BV), the forward overall variance (FV) and the difference between these two (DV = BV – FV). FV and BV are calculated as the sum of the var* for each of the three acceleration axes, on the sequences immediately before (BV) and after (FV) a potential beginning or end of a rally (the size of the sequences was subject to optimization). To be able to truly detect the best point describing the beginning or end of each rally, we also calculated peaks on the DV feature. Figure 3: The final front showing the best solutions based on the Calculating the peaks was done the same way as for the shot classification error and the number of shots outside of detected detection. The minimum distance between peaks was subject to active play. the optimization. Since our objective was to accurately detect the duration of the rallies, we chose one solution from the middle of the front and for 4.2 Problem Formulation this solution, we calculated the distribution of the durations of the For each data point we calculated the previously described rallies. The comparison with the manually labeled rallies can be features, and set a rule for detecting the beginning of a rally and a seen in Figure 4. rule for detecting the end of a rally. A data point is marked as the beginning of a rally if it satisfies the following rule: . A data point is marked as the end of the rally if it satisfies the following rule: , where parameters p 1, p 2, p 3, p 4 were determined through optimization. Both rules consist of two parts. The first part Figure 4: Comparing manually labeled (left) and automatically determines the threshold for the change in activity before and after detected distributions for play durations. a potential beginning or end of a rally. For the beginning of a rally, this difference is usually larger because a rally often starts We can clearly see the similarities between the distributions. The explosively and ends gradually, so the thresholds p 1 and p 2 can reason for the detected distribution having more very short rallies be different. The second part is the same for both rules and serves is that the algorithm detects even small starts of movement that we to remove false detections due to the variation in intensity during did not label as rallies because they were too short. For example, a the rally by specifying that the activity, either before or after the server hitting the net with the first serve results in the returner beginning or end of a rally, should be low. making just a small movement. So altogether we optimized six input parameters: sequence size, By combining the classified shot types, detected active playing minimum distance between peaks, p 1, p 2, p 3 and p 4. phases and locations from the GPS, we can calculate several useful metrics that help remove subjectivity from the game and allow for objective evaluation of different tactical approaches and training 4.3 Experimental Setup routines. An example of such a view can be seen in Figure 5, where To optimize the two objectives – classification error and the we present the heat map of a player’s position during active play and number of shots not inside the detected rallies – we used the well- combine it with forehand and backhand shots as points of different known evolutionary multi-objective optimization algorithm called color and size. We also included a dashed line that separates part of NSGA-II [4]. The population size was set to 25, the stopping the court where more backhands are played from the one where criterion was set to 10,000 solution evaluations, and the more forehands were played. We can see that the player played more tournament selection was used. aggressively on the left side, thus his heat map is closer to the baseline. On the right side, a less aggressive approach allowed him 4.4 Results to play more forehands and thus he dictated play by playing more The final front of the optimization can be seen in Figure 3. We often with his better shot. can see a typical result for a multi-objective optimization problem, a non-dominated front showing a tradeoff between objectives. We can also see a knee on the front labeled with a circle. In this solution six shots are missed, since they occurred without the surrounding intense activity which accompanied other shots. An example is a player hitting the ball out of court after the rally 45 6. ACKNOWLEDGMENTS We would like to thank Catapult Sports for their sound advice and cooperation, and the national research agency (ARRS) for providing the research grant under the project PR-06884. 7. REFERENCES [1] Renò, Vito, et al. "A technology platform for automatic high- level tennis game analysis." Computer Vision and Image Understanding (2017). Figure 5: Heat map of a player’s position during active play [2] Liaw, Andy, et al. "Classification and regression by combined with shot locations (blue = forehand, RandomForest." R news 2.3 (2002): 18-22. red = backhand) and their Peak_strength (size of points). [3] Liberty, Edo, Kevin Lang, and Konstantin Shmakov. "Stratified sampling meets machine learning." International Conference on Machine Learning. 2016. 5. CONCLUSION [4] Deb, Kalyanmoy, et al. "A fast and elitist multi-objective In this article we presented a two-part algorithm for analyzing genetic algorithm: NSGA-II." IEEE transactions on wearable sensor data for professional tennis players. Firstly, we evolutionary computation 6.2 (2002): 182-197. detected and classified different shot types, and secondly, we distinguished the active playing phases from the time in-between [5] Yu, Xinguo, et al. "A trajectory-based ball detection and points. By combining the procedures, players can get a unique tracking algorithm in broadcast tennis video." Image perspective on their game which enables objective analysis in the Processing, 2004. ICIP'04. 2004 International Conference tactical and physical sense. on. Vol. 2. IEEE, 2004. For the future, we plan to equip both players with the same type [6] Pei, Weiping, et al. "An embedded 6-axis sensor based of sensor, and by measuring the time difference between their recognition for tennis stroke." IEEE International Conference shots and by calculating the distance between them, we will be on Consumer Electronics (ICCE), 2017. able to calculate the average speed of ball and thus additionally [7] Yang, Disheng, et al. "TennisMaster: an IMU-based online quantify the quality of each shot. serve performance evaluation system." Proceedings of the 8th Augmented Human International Conference. ACM, 2017 46 Anytime Benchmarking of Budget-Dependent Algorithms with the COCO Platform Tea Tušar Nikolaus Hansen Dimo Brockhoff Department of Intelligent Inria Inria Systems École Polytechnique CMAP École Polytechnique CMAP Jožef Stefan Institute (UMR 7641) (UMR 7641) Ljubljana, Slovenia Palaiseau, France Palaiseau, France tea.tusar@ijs.si nikolaus.hansen@inria.fr dimo.brockhoff@inria.fr ABSTRACT the data profile has another important advantage—it can Anytime performance assessment of black-box optimization be used to represent algorithm performance aggregated over algorithms assumes that the performance of an algorithm at multiple runs on different problems of the same dimension- a specific time does not depend on the total budget of func- ality (see Section 2 for more details). This considerably al- tion evaluations at its disposal. It therefore should not be leviates presentation and understanding of algorithm results used for benchmarking budget-depending algorithms, i.e., on a large number of problems. algorithms whose performance depends on the total budget of function evaluations, such as some surrogate-assisted or The underlying assumption in anytime performance assess- hybrid algorithms. This paper presents an anytime bench- ment is that the performance of an algorithm at a specific marking approach suited for budget-depending algorithms. runtime does not depend on the total budget of function The approach is illustrated on a budget-dependent variant evaluations. That is, performance of an algorithm at 1000 of the Differential Evolution algorithm. function evaluations is expected to be the same if the al- gorithm was run with a budget of 1000 or 100 000 function evaluations (everything else being equal). If this is not the 1. INTRODUCTION case, data profiles should not be employed to infer perfor- In black-box optimization, the problem to be optimized can- mance of the same algorithm with a total budget different not be explicitly written as a function of its input parame- from the one used in the experiments. ters (if an underlying function exists, it is unknown). This is often the case with real-world problems where solutions Algorithms can depend on the total budget for different rea- are evaluated using simulations. Without the possibility of sons. Consider for example surrogate-assisted approaches. exploiting the structure of the function, optimization algo- They construct surrogate models of the optimization prob- rithms resort to repeatedly sample its decision space and lem and combine actual function evaluations with evalua- use previously evaluated solutions to steer the search to- tions on the models. While some algorithms work in a wards promising regions. Since the evaluations of real-world budget-independent way, e.g. [6], others save some true func- problem functions are often more time-consuming than the tion evaluations for the end (just before the budget is ex- internal computations of optimization algorithms, the run- hausted), making them budget-dependent, e.g. [9]. Simi- ning time, or runtime, of an algorithm is generally measured larly, hybrid genetic algorithms that combine genetic algo- by counting the number of performed function evaluations. rithms with local search methods can reserve a number of fi- The goal of an algorithm in black-box optimization is thus nal function evaluations to additionally improve the current to find satisfactory solutions to the given problem in as few best solutions [2]. Another example of budget-dependent al- function evaluations as possible. gorithms are evolutionary algorithms that set any of their parameters based on the total budget [1]. When measuring the performance of an algorithm in the black-box setting, we are interested in the required runtime To address this issue, we propose an approach for bench- to reach a target value. Or rather, we wish to obtain all marking budget-dependent algorithms that allows anytime runtimes corresponding to increasingly difficult targets [3]. performance assessment of their results. It is based on the In problems with a single objective, the targets are usu- anytime benchmarking from the Comparing Continuous Op- ally defined as differences from the optimal function value, timizers (COCO) platform [4]. The approach is demon- while in problems with multiple objectives, the targets are strated on a budget-dependent variant of Differential Evo- determined as differences from the optimal value of a mul- lution (DE) [8] ran on the bbob test suite [5]. tiobjective performance indicator. In the following, we first present some background on any- The proportion of reached targets plotted against the run- time benchmarking with the COCO platform (Section 2). times in the sense of an empirical cumulative distribution The new approach is described in Section 3 followed by a function yields a data profile [7]—a graph showing the any- discussion on its time complexity. An illustration with the time performance of an algorithm, essentially mimicking the DE algorithm is shown in Section 4. Section 5 presents some convergence graph (the plot of best found function or indica- concluding remarks. tor values over time). In addition to being easy to interpret, 47 2. ANYTIME BENCHMARKING IN COCO analogous way as before, but this time the y axis shows the COCO (https://github.com/numbbo/coco) is a platform proportion of targets reached out of N · l ones. that facilitates benchmarking of optimization algorithms by automatizing this procedure and providing data of previ- Finally, data profiles are also able to aggregate runtime re- ously run algorithms for comparison [4]. An important part sults over problems of the same dimensionality that opti- of COCO’s anytime benchmarking approach [3] is the pre- mize a different function. Imagine a test suite consisting sentation of algorithm results in the form of data profiles [7]. of m such problems with multiple instances. After boot- strapping is performed for each problem separately, there Consider a single run of algorithm A on problem p. Given l are m · N · l function and target pairs and the same number increasingly difficult targets τ1, τ2, . . . , τl, it is easy to com- of bootstrapped runtimes. The aggregated data profile for pute the corresponding runtimes rp, rp, . . . , rp needed by algorithm A can thus be constructed by plotting for each 1 2 l algorithm A to reach each of these targets. If the target τj number of evaluations the proportion of function and target was not reached, rp is undefined. A data profile for algo- pairs reached by A in a runtime equal to or smaller than the j rithm A is then constructed by plotting for each number of number of evaluations. evaluations the proportion of targets reached by A in a run- time equal to or smaller than the number of evaluations. In It is important to note that runtimes are never aggregated other words, a data profile is the empirical cumulative dis- over different dimensions since problem dimension is often tribution function of the recorded runtimes rp, rp, . . . , rp. used as an algorithm parameter. This also allows scalability 1 2 l studies. See [3] for more details on COCO’s performance Data profiles can be further exploited to show aggregated in- assessment procedure. formation over randomized repetitions of running algorithm A on problem p. Instead of using repeated runs of A on p 3. A BENCHMARKING APPROACH FOR (which is sensible only for stochastic algorithms), random- BUDGET-DEPENDENT ALGORITHMS ization in COCO is achieved by running the same algorithm A on different instances of problem p (for example, trans- The idea for benchmarking a budget-dependent algorithm A lated versions of the same problem). is very simple: the algorithm is run with increasing budgets and the resulting runtimes are presented in a single data Consider k instances of the problem p, denoted here as p(θ profile. This is achieved by means of an ‘artificial’ algorithm 1), p(θ e A that works as follows. 2), . . . , p(θk ). Like before, the runtime rp(θi) at which al- j gorithm A achieves target τj on problem instance p(θi) can be easily calculated for each i and j and is undefined when Consider K increasing budgets b1, b2, . . . , bK and K budget- the target has not been reached. In order to be able to com- dependent algorithm variants Ab , A , . . . , A . The algo- 1 b2 bK pare algorithms of different success probabilities (for exam- rithm e A first works as algorithm Ab for budgets b ≤ b 1 1, then ple an algorithm that always reaches difficult targets, but works as algorithm Ab for budgets b, where b 2 1 < b ≤ b2, does this slowly, with an algorithm that sometimes reaches and so on, finishing by mimicking algorithm Ab for budgets K a target quickly while other times fails to reach it at all), we b, where bK−1 < b ≤ bK (see also Figure 1). In an algorith- simulate restarts of each algorithm via a bootstrapping pro- mic notation (where xi denotes the ith solution explored by cedure. The N bootstrapped simulated runtimes r the corresponding algorithm): j,1, rj,2, . . . , rj,N of the artificially restarted algorithm to reach a target τ b j are computed from the recorded runtimes rp(θi) of 0 ← 0 . Initialize a budget preceding b1 j algorithm A (for a large N, e.g. N = 1000) as: for j ← 1, . . . , K do . Iterate over budgets for i ← 1, . . . , bj−1 do A for c ← 1, . . . , N do . Repeat N times b (x ignoring its output j i) . Run Abj r end for j,c ← 0 . Initialize runtime loop for i ← bj−1 + 1, . . . , bj do i ← random({1, . . . , k}) . Choose an instance e A(xi) ← Ab (x j i) . e A mimics Abj end for if rp(θi) is defined then . Successful j end for rj,c ← rj,c + rp(θi) j break loop Although the first b are else . Unsuccessful j−1 evaluations of the algorithm Abj r ignored by e A, they need to be performed so that Ab is in j,c ← rj,c + rp(θi) max j end if the correct state at evaluation bj−1 + 1, when it starts to be end loop mimicked by algorithm e A. end for return r As shown with the red run in Figure 1, A might not con- j,1, rj,2, . . . , rj,N bj tribute to e A at all if the performance of Ab is better for i if at least one of the recorded runtimes is finite. Note that some bi < bj . On the other hand, if Ab is significantly bet- j the total runtime of A on p(θ ter than A (for example, the green vs. the yellow run), i), rp(θi) max , is added each time an bj−1 unsuccessful trial is picked. Runtimes rj,1, rj,2, . . . , rj,N are this causes a ‘jump’ in the performance of e A. Note also undefined for targets τj that were not reached in any of the that the best-so-far profile of e A does not necessarily follow problem instances. The resulting N · l runtimes (of which the overall best-so-far profile of Ab , but only its best-so-far j some undefined) are used to construct the data profile in an profile after bj−1 function evaluations (notice the yellow and 48 20 Table 1: An upper bound of function evaluations 40 required for benchmarking budget-dependent algo- 60 rithms with K budgets between 1 and 10M that are 80 equidistant in the logarithmic scale, for some se- 100 lected values of K and M . l m 101/K+M −1 M 101/K −1 3 4 5 10 4.859 48.618 486.208 (to be minimized) 20 9.188 91.947 919.540 K 50 22.198 222.165 2.221.835 100 43.889 439.270 4.393.094 Function or quality indicator value 0 20 40 60 80 100 Table 2: Population size of the budget-dependent Number of function evaluations DE computed for some selected values of budget multipliers mbudg and problem dimensions n. Figure 1: An illustration of the ‘artificial’ algorithm m 3 log2 (n · m budg 10 budg) 10 100 1000 e A constructed from five runs of algorithm A (Ab means the algorithm was run with the budget of 2 5 15 32 b function evaluations). Thin and thick lines show 5 8 21 41 n the actual and best-so-far performance for each run 10 12 27 48 of A, respectively. 20 15 32 55 black lines after 40 function evaluations). budget-independent, a study shows that setting its param- eters, especially population size, in connection to the total Composing the performances of algorithm A with different budget of evaluations can improve its results [1]. budgets into algorithm e A results in an estimation of the anytime performance of A. The quality of the estimation In the experiments we use the DE implementation from the depends on the number of budgets K—more budgets en- scipy Python package (https://www.scipy.org/) with the able a better estimation, but make the procedure more time following parameters: consuming. – Initialization = Latin Hypercube sampling One could run the budget-dependent variants of algorithm – DE strategy = best/1/bin A for every budget between 1 and b – Population size = variable (see text) K thus obtaining the best possible estimate. However, this would require – Weight F = random in the interval [0.5, 1) – Crossover probability CR = 0.7 bK X bK (bK + 1) – No local optimization of the final solution j = 2 – Relative tolerance for convergence = 10−9 j=1 total evaluations. Diluting the budgets by taking only every M th one does not help to significantly reduce the number This implementation computes the population size based of total evaluations. A more promising approach is that on the problem dimension n (for a user-specified multiplier of using equidistant budgets in the logarithmic scale. For mpop, the population size is calculated as n·mpop). This was example, K such budgets between 1 and 10M require bypassed in order to make population size budget-dependent. For a problem with n dimensions and a budget multiplier M K X 101/K+M − 1 m 10j/K = budg, the actual budget in COCO is computed as n · mbudg 101/K − 1 and the population size of DE is calculated as j=0 total evaluations. Table 1 contains total evaluations for this 3 log2 (n · m 10 budg) . case for some values of K and M . The actual number of Table 2 gathers the values of this formula for selected budget evaluations is likely to be smaller than these numbers due multipliers mbudg and problem dimensions n. to some consecutive (small) budgets being rounded to the same integer number. The experiment consisted of running five instances of DE with budget multipliers from {10, 31, 100, 316, 1000} and at 4. EXAMPLE the same time composing their performances into the any- We present a small example to demonstrate the proposed time artificial algorithm called ‘DE-anytime’. All algorithms anytime benchmarking procedure on the COCO platform. were run on the 24 problems functions from the bbob test The algorithm used in this example is a budget-dependent suite [5] with dimensions n in {2, 3, 5, 10, 20}. Each problem variant of Differential Evolution (DE) [8], a well-known evo- was instantiated 15 times. The results for dimension 10 are lutionary algorithm. While the original DE algorithm is presented in Figure 2. In data profiles plots produced by 49 1.0 bbob f1-f24, 10-D best 2009 31 targets RLs/dim: 0.5. 50 sessment of their results. The approach demands repeated from refalgs/best2009-bbob.tar.gz 15 instances runs of an algorithm with increasing budgets. Depending on 0.8 DE-anytim the number and size of these budgets, it can take a signifi- cant amount of time (it is quadratic in the maximal budget DE-1000 in the worst case). By using budgets that are equidistant in 0.6 the logarithmic scale, the time complexity depends linearly DE-316 on the maximal budget, making the approach more usable in practice. An example experiment showing how to use this 0.4 DE-100 approach in COCO will be available in COCO v2.2. 0.2 6. ACKNOWLEDGMENTS DE-31 The authors wish to thank Joshua Knowles for helpful dis- Fraction of function,target pairs cussions on this topic during the SAMCO workshop (Work- 0.0 v0.0.0 DE-10 shop on Surrogate-Assisted Multicriteria Optimization) at 0 1 2 3 the Lorenz Center in Leiden, The Netherlands, 2016. log10(# f-evals / dimension) The authors acknowledge the financial support from the Slovenian Research Agency (project No. Z2-8177). This Figure 2: Data profiles for budget-dependent vari- work is part of a project that has received funding from ants of DE run with budgets of {10, 31, 100, 316, 1000}·n the European Union’s Horizon 2020 research and innovation number of evaluations and the ‘artificial’ algorithm program under grant agreement No. 692286. constructed from these variants estimating anytime performance of DE. See text for further information. This work was furthermore supported by a public grant as part of the Investissement d’avenir project, reference ANR- 11-LABX-0056-LMH, LabEx LMH, in a joint call with Gas- COCO, the function evaluations are always divided by the pard Monge Program for optimization, operations research problem dimension n and shown on a logarithmic scale. and their interactions with data sciences. The benchmark setting used in this example is COCO’s ex- 7. REFERENCES pensive setting, in which the number of evaluations is lim- [1] A. S. D. Dymond, A. P. Engelbrecht, and P. S. Heyns. ited to 1.000n and the 31 targets are defined in a relative The sensitivity of single objective optimization way—according to the performance of a virtual algorithm algorithm control parameter values under different denoted with ‘best 2009’ that is comprised of the best re- computational constraints. In Proceedings of CEC 2011, sults achieved by 31 algorithms at the Black-box Optimiza- pages 1412–1419, 2011. tion Benchmarking (BBOB) workshop in 2009. The targets [2] T. A. El-Mihoub, A. A. Hopgood, L. Nolle, and are chosen from [10−8, ∞) such that the ‘best 2009’ algo- A. Battersby. Hybrid genetic algorithms: A review. rithm just failed to reach them within the given budget of Engineering Letters, 13:124–137, 2006. nmbudg evaluations, with 31 different values of mbudg chosen equidistantly in logarithmic scale between 0.5 and 50. [3] N. Hansen, A. Auger, D. Brockhoff, D. Tušar, and T. Tušar. COCO: Performance assessment. ArXiv We are showing the results for dimension 10, since the differ- e-prints, arXiv:1605.03560, 2016. ences among DE instances are best visible for this dimension. [4] N. Hansen, A. Auger, O. Mersmann, T. Tušar, and Note that the algorithms were stopped at the moment de- D. Brockhoff. COCO: A platform for comparing noted by the large cross, but the data profiles increase also continuous optimizers in a black-box setting. ArXiv beyond that point due to bootstrapping (see Section 2). e-prints, arXiv:1603.08785, 2016. [5] N. Hansen, S. Finck, R. Ros, and A. Auger. From Figure 2 we can observe that DE variants with a bud- Real-parameter black-box optimization benchmarking get of 10n and 31n evaluations achieve a very similar perfor- 2009: Noiseless functions definitions. Research Report mance in the first 10n evaluations. Other DE variants are RR-6829, INRIA, 2009. noticeably different from the first two and also among them- [6] D. R. Jones, M. Schonlau, and W. J. Welch. Efficient selves, with those with lower budgets converging faster at the global optimization of expensive black-box functions. beginning of the run. This confirms the findings from [1] that Journal of Global Optimization, 13(4):455–492, 1998. better performance can be achieved by fitting the population [7] J. J. Moré and S. M. Wild. Benchmarking size to the total budget. derivative-free optimization algorithms. SIAM Journal on Optimization, 20(1):172–191, 2009. The dark blue data profile corresponding to the ‘DE-anytime’ [8] R. Storn and K. Price. Differential evolution – A simple artificial algorithm follows the five underlying algorithms as and efficient heuristic for global optimization over expected. The accuracy of this estimate could be further continuous spaces. Journal of Global Optimization, improved if a higher number of different budgets was used. 11(4):341–359, 1997. [9] V. Volz, G. Rudolph, and B. Naujoks. 5. CONCLUSIONS Surrogate-assisted partial order-based evolutionary We presented a novel approach for benchmarking budget- optimisation. In Proceedings of EMO 2017, pages dependent algorithms that enables anytime performance as- 639–653, 2017. 50 Criteria for Co-existence of GM and Conventional Maize Production Marko Debeljak Florence Leprince Sašo Džeroski Aneta Trajanov Jožef Stefan Institute and ARVALIS-Institut du vegetal, Jožef Stefan Institute and Jožef Stefan Institute and Jožef Stefan International 21, chemin de Pau, Jožef Stefan International Jožef Stefan International Postgraduate School Montardon, France Postgraduate School Postgraduate School Jamova 39, Ljubljana, +33 5 59 12 67 79 Jamova 39, Ljubljana, Jamova 39, Ljubljana, Slovenia f.leprince@arvalisinstitu Slovenia Slovenia +386 1 477 3124 tduvegetal.fr +386 1 477 3217 +386 1 477 3662 marko.debeljak@ijs.si saso.dzeroski@ijs.si aneta.trajanov@ijs.si ABSTRACT assess the co-existence performance of applied management The criteria for co-existence of genetically-modified (GM) and strategies in GM fields, and ii) to identify efficient crop conventional (non-GM) crops must reflect the best available management measures that enable the co-existence of GM and scientific evidence on mixture between these two types of crops. conventional crop production systems. Co-existence strategies based on fixed isolation distances are not The identification of co-existence criteria is currently based on in line with the EC guidelines on co-existence, which require two approaches. The first uses a mechanistic matrix modeling criteria adaptable to local constraints. In this paper, we apply approach that is based on a theoretical description of pollen data mining for identification of co-existence criteria of maize dispersal, while data from field experiments are used for production. We use classification trees to generate co-existence calibrations and validations of such models [1]. The second criteria for GM and conventional maize fields. The data used in approach is based on empirical knowledge about co-existence, this study were provided by ARVALIS and consisted of several obtained mostly from observations and experiences from surveys of outcrossing between pairs of maize fields. Based on growing GM crops under real production conditions, where the the model structure, the most important co-existence criteria are performance of fixed co-existence measures is used and flowering time lag, wind direction, presence of isolation rows evaluated. Such an empirical model, called a global index, has and distance between the GM and conventional field. The co- been developed by [11]. existence criteria generated from the model for prediction of outcrossing were applied on an independent Spanish dataset. In this study, we propose a third approach that employs The results are meaningful and in accordance with literature and techniques of data mining to real data about cross-pollination have high potential for application in the development of between GM and conventional crops grown under different crop computer based co-existence decision support system. management practices. Our goal is to identify co-existence criteria about the adventitious presence of GM maize in the conventional maize production, using the official threshold level Keywords of 0.9%, from the structure of the induced predictive models Genetically modified crops, GM maize, co-existence criteria, built from real data. classification trees, random forests. 2. MATERIALS AND METHODS 1. INTRODUCTION 2.1 Data Co-existence is concerned with the potential economic impact of the mixture of genetically modified (GM) and non-GM crops, In this study, we used data provided by ARVALIS - Institut du végétal, France, and Institut de Recerca i Tecnologia the identification of workable management measures to Agroalimentàries (IRTA) minimize mixture, and the cost of these measures [5]. Co- , Spain. The data provided by existence applies only to approved GM crops that were ARVALIS were used for construction of the data mining models considered to be safe prior to their commercial release and and the Spanish data were used for validation of the induced co- safety issues fell outside its remit [14]. EC regulation 1829/2003 existence criteria. (article 43) [6] provides guidelines to develop national or The data provided by ARVALIS were from surveys of regional strategies and best practices to ensure co-existence. outcrossing (gene flow from GM donor crop to recipient non- However, the selection of preventive co-existence criteria is the GM crop) between pairs of 88 maize fields in the Pau and individual responsibility of each Member State. Toulouse regions in the South – West of France, in the period The level of purity needed to ensure co-existence is defined by a from 2001-2007. Each field was described with the following tolerance threshold. The EU accepts an adventitious or variables: location of the fields, distances between donor and technically unavoidable presence of authorized GM material in recipient fields, locations of sampling points, number of border non-GM food and feed up to 0.9% and the main task of co- rows, field size, sowing date, flowering date, flowering time-lag existence is to find out by which means the adventitious between donor and recipient fields, isolation distance between presence can be kept below the accepted threshold level. In pairs of donor and recipient fields, prevailing wind direction particular, the prediction of adventitious presence of GM from donor fields during the flowering period, and the material in neighboring non GM fields is required in order i) to percentage of outcrossing (outcrossing rate) in recipient fields using real-time quantification system-polymerase chain reaction. 51 The Spanish data were collected at harvesting time on 13 non- presence of GM material, while only around 10% were above GM maize fields in 400 ha large maize crop area Pla de Foixà this threshold. In cases like this, the accuracy is not the optimal region (Girona), in Catalonia, Spain, in 2004, 2005 and 2006. performance metric for evaluation of the data mining models. During these years, the crop type, variety, sowing and flowering Therefore, we used an additional performance metric - the Area dates of maize fields were recorded, as well as meteorological Under the Receiver-Operator characteristic Curve (AUROC) data. At harvesting time, samples were collected on non-GM [7], to more objectively evaluate the performance of the models. maize fields (7 fields in 2004, 4 fields in 2005 and 4 fields in AUROC is defined for binary classification, where one of the 2006). The samples were analyzed by RTQ-PCR to evaluate the classes is considered positive. A discrete classifier produces a cross-pollination between GM and conventional maize [11]. The pair of False Positives Rate (FPR: negatives incorrectly spatial distribution of crops in the selected region was described classified / total negatives) and True Positives Rate (TPR: by maps generated from aerial photographs. positives correctly classified / total positives), which corresponds to a single point in the ROC space, while classifiers that return a probability value for the positive class correspond to an ROC curve. AUROC values of 0.7 and higher are considered to indicate a good fit to the data [7]. 2.3 Data preprocessing We used outcrossing grains as a unit of the outcrossing rate for all samples and surveys given in percentage of DNA [10]. From the original data described in section 2.1, we calculated the following attributes that we later used in the data mining analyses: - Minimal distance from a sampling point to the donor field [m] - Isolation distance (minimal distance between the donor and recipient field) [m] Figure 1: Example of the field-to-field map where the donor maize field is on the left and the recipient maize field is on - Presence of isolation rows [yes, no] the right side. Sampling points in the recipient maize field - Wind direction [0 – upwind (from recipient field), 1 – are indicated with lines. downwind (to recipient field)] 2.2 Data mining methods - Flowering time-lag [minimal (0-7 days), medium (8-14 days) To find interactions between the attributes describing and large (more than 15 days)] geographical, environmental and management parameters and - Common border length between donor and recipient field [m] outcrossing rates measured at sampled fields, we used data mining methods for induction of decision trees, which are - Outcrossing rate [1 - if < 0.9% GM grains, 2 - if ≥ 0.9% GM ideally suited for analysis of complex ecological data. Decision grains] trees predict the value of a dependent variable (target) from a set We discretized some of the initial attributes in order to obtain of independent variables (attributes). In our case, the dependent comprehensible and easily interpretable predictive models of (target) variable is the outcrossing rate between a GM and a outcrossing. The ranking and threshold values for Flowering conventional maize in the field, which can have two values: 1 time-lag attribute were selected according to expert knowledge (below the threshold of 0.9% of adventitious presence) or 2 about maize production, while the target variable Outcrossing (above the threshold of 0.9% of adventitious presence). In this rate was discretized according to the accepted European study, we used classification trees to develop predictive models threshold (0.9 % grains). for co-existence of GM maize production. To deal with the high imbalance in the data, we applied To evaluate the induced data mining models, we used two methods, such as up-sampling and down-sampling of the measures of performance or agreement of the discrepancy dataset, in order to create a more balanced dataset. However, the between measurements and predictions. We first calculated newly obtained balanced dataset did not improve the results, so classifier’s accuracy as the proportion of samples for which the we stayed with the original dataset. category (below threshold vs. above threshold) was correctly predicted. We used 10-fold cross-validation as the most The Spanish data had a different structure than the French common and standard way of estimating the performance dataset, so to use it for validation of the generated co-existence (accuracy) of a learning technique on unseen cases [15]. criteria, we standardized the data to achieve the same dataset structure as in the case of the ARVALIS dataset used for Data from real field studies describe the actual practices applied building data mining models (Section 3.1). First, we unified the on the fields, where in the case of growing GM crops, units of the measurements on the fields. Then, we reorganized incorporation of precautions for prevention of outcrossing of and calculated the data for pair-based comparisons of donor- GM material to conventional fields is obligatory. Therefore, recipient fields to create a setting similar to the setting in the most often, these data are highly imbalanced, having a low ARVALIS experiments. number of outcrossing events. In our case, most of the samples (around 90%) were below the threshold of 0.9% of adventitious 52 3. RESULTS The co-existence criteria (Table 2) derived from the classification tree generated from the French data, were 3.1 Data mining models confirmed by the Spanish data in 88.6% cases. Validation of the individual criterions (rules) has shown that the criterion 1 (time- To generate classification trees, we used the algorithm J4.8, an lag) was confirmed in 90% of the cases, 76% of cases confirmed implementation of the C4.5 algorithm within the WEKA suite criterion 2 (combination of time-lag and wind direction), while [15]. criterion 4 was valid for 100% of the cases. None of the surveyed Spanish maize fields had protection rows, therefore, The classification model constructed on training data (Fig. 2) we were not able to validate criterion 3. correctly classified 94.31% of the instances, while the cross- validated model correctly classified 90.24% of the instances The validation of the outcrossing model (Figure 2) and the co- (Table 1). However, the accuracy is sensitive to imbalanced data existence criteria, confirmed that there is a high potential for and therefore, the model correctly classifies most of the their application in the assessment process of co-existence instances belonging to class 1, but it misclassifies most of the issues. instances belonging to class 2. The most unbiased measure of the goodness of a model is its AUROC value. The AUROC values of the classification trees obtained on training data and with cross-validation are given in Table 1. The AUROC value of the classification tree obtained with cross-validation is (0.576). This value is very close to 0.5 (the diagonal y = x), which means that the predictive power of the classification tree obtained with cross-validation is not very high. The predictive power of the classification tree obtained on training data according to its AUROC value (0.850) is much higher and indicates good predictive power. Table 1: Accuracy and AUROC values for the J48 algorithm applied to the ARVALIS data. Accuracy AUROC value J48 on training data 94.31% 0.850 J48 10-fold cross- 90.24% 0.576 validation Figure 2: Outcrossing model induced by J48 (values in the 3.2 Co-existence criteria leaves: 1: outcrossing < 0.9% grains, 2: outcrossing ≥ 0.9% grains). The predictive power of the obtained classification tree (Fig. 2) is not very high due to the imbalanced data, therefore it cannot Table 2: Proposed co-existence criteria for GM and be used for making predictions about the outcrossing between conventional maize production. GM and non-GM fields. However, its descriptive power is very 1. If the flowering time-lag is ≥ 15 days, then the relevant. The topmost part of the classification tree includes the outcrossing rate in a recipient field is below 0.9%. attributes time-lag, wind direction, isolation distance and presence of isolation rows. Because of their dominant position 2. If the flowering time-lag is less than 15 days and if in the model structure, they can be recognized as the ones which the prevailing wind direction during flowering play the most important role in the cross-pollination process and period is from recipient to donor field, the as such they could present the most important criteria for co- outcrossing rate of a recipient field is below 0.9%. existence of GM and conventional maize production. 3. If the time-lag is less than 15 days and the wind Experts from ARVALIS and IRTA, as well as extensive direction during flowering period is from donor to literature [4,9,11,12,13] confirm the importance of the attributes recipient field, but the donor or the recipient filed captured in the model structure for the outcrossing process. has an isolation or protection row, then the Therefore, we focused on the parts of the model that correctly outcrossing rate in the recipient field is below 0.9%. predicts outcrossing below 0.9% (leaves in the tree that provide most accurate predictions, presented by the number of examples 4. If the time-lag is less than 15 days and wind that fall in a leaf and the number of incorrectly classified direction during flowering days is from donor to examples in Figure 2) and arranged our findings into a coherent recipient field, but there are no isolation nor and consistent list of co-existence criteria (Table 2). protection rows and the distance between the donor and the recipient field is more than 50 m, then the In order to obtain robust and applicable co-existence criteria, we outcrossing rate in the recipient field is below 0.9%. validated them on an independent dataset provided by IRTA, Spain. 53 4. CONCLUSIONS of transgenes in maize between neighboring maize fields. In: Proceedings of the 19th International Conference Compared to most of the studies about maize cross-fertilization Informatics for Environmental Protection (EnviroInfo) due to pollen flow, we used the opportunity to gain new (Hrebicek J., Jaroslav R. eds.). Brno, Czech Republic. pp. knowledge about this phenomenon by applying data mining 610–614. techniques to explore the information stored in datasets of real [4] Della Porta, G., Ederle, D., Bucchini, L., Prandi, M., maize growing management. This allowed us to overpass some Verderio, A., Pozzi, C. 2008. Maize pollen mediated gene shortages of previous field experimental designs that were flow in the Po valley (Italy): source-recipient distance and mostly oriented toward worst-case scenarios (e.g., small donor effect of flowering time. Eur J Agron 28, 255–265. field placed in the center of large recipient field [2,3] or spatial arrangements and distribution of donor and recipient [5] European Commission, 2003a. Commission experimental fields that were too simplified compared to the real recommendation on guidelines for the development of ones [8]. national strategies and best practices to ensure the coexistence of genetically modified crops with Data describing real field experiments that involve GM crops conventional and organic farming. Off J Eur Union L 189, are in general very imbalanced, due to the fact that farmers are 36-47. obliged to take measures to prevent or minimize the outcrossing from GM to conventional fields. In addition, field experiments [6] European Commission, 2003b. EC regulation no. about outcrossing rates show a fast decrease of outcrossing by a 1829/2003 of the European Parliament and of the council distance from donor field [13]. Because of these reasons, the of 22 September 2003 on genetically modified food and datasets we used contained much larger number of sampling feed. Off J Eur Union L 268, 1-23. points with outcrossing below 0.9%, which resulted in a very [7] Fawcett, T. 2006. An introduction to ROC analysis. imbalanced structure of the data. Pattern Recogn Lett 27, 861-874. To mitigate this problem we studied different performance [8] Loos, C., Seppelt, R., Meier-Bethke, S., Schiemann, J., measures to assess the goodness of the obtained models. These Richter, O. 2003. Spatially explicit modelling of transgenic measures showed that the models are very precise in predicting maize pollen dispersal and cross-pollination. J Theor Biol the situations when outcrossing is less than 0.9%, but not that 225, 241–255. precise when predicting the situations when the outcrossing is [9] Meier-Bethke, S., Schiemann, J. 2003. Effect of varying above 0.9%. Therefore, we accepted a compromise to make a distances and intervening maize fields on outcrossing rates recommendation about the co-existence criteria that are taken of transgenic maize. In: 1st European Conference on the from the structure of the predictive outcrossing model. However, Co-existence of Genetically Modified Crops with the part of the model describing conditions, which lead to the Conventional and Organic Crops (Boelt B. ed.) . Research outcrossing rate above 0.9% is not very reliable. Center Flakkebjerg, pp. 77–78. Our study made a significant progress about using data mining [10] Messéan, A., Angevin, F., Gomez-Barbero, M., Menrad, methods for identification of co-existence criteria. However, the K., Rodriguez Cerezo, E., 2006. New case studies on the imbalance of the data is a problem that needs to be addressed by coexistence of GM and non-GM crops in European applying data mining methods that are suited for analyzing that Agriculture. Technical Report Series of the Joint Research kind of data, such as cost-sensitive learning. Finally, in this Center of the European Commission. Institute for study, we were focused on the pair-based effects of outcrossing Prospective Technological Studies, Sevilla. between GM and non-GM fields. To assess the multi-field [11] Messeguer, J., Peńas, G., Ballester, J., Bas, M., Serra, J., effects on the outcrossing rate at a selected recipient field, a Salvia, J., Palaudelmàs, M., Melé, E. 2006. Pollen- different data mining setting should be created and other data mediated gene flow in maize in real situations of mining methods can be exploited, such as methods for multi- coexistence. Plant Biotechnol J 4, 633–645. target prediction and inductive logic programming. Furthermore, this study shows the ability of data mining methods to extract [12] Palaudelmàs, M., Messeguer, J., Peñas, G., Serra, J., useful information about co-existence issues from data Salvia, J., Pla, M., Nadal, A., Melé, E. 2007. Effect of describing real and not experimental maize production settings. sowing and flowering dates on maize gene flow. In: Book By that, data mining models for prediction of outcrossing under of abstracts of the third International Conference on real crop production conditions could be successfully Coexistence between Genetically Modified (GM) and non- incorporated in a computer based co-existence decision support GM-based Agricultural Supply Chains (Stein A.J., system. Rodríguez-Cerezo E. eds.). EC, pp. 235–236. [13] Sanvido, O., Widmer, F., Winzlerm, M., Streitm, B., 5. REFERENCES Szerencsits, E., Bigler, F. 2008. Definition and feasibility [1] Beckie, H.J., Hall, L.M. 2008. Simple to complex: of isolation distances for transgenic maize. Transgenic Res Modelling crop pollen-mediated gene flow. Plant Sci 175, 17, 317-355. 615-628. [14] Schiemann, J. 2003. Co-existence of genetically modified [2] Belcher, K., Nolan, J., Phillips, P.W.B. 2005. Genetically crops with conventional and organic farming. Environ modified crops and agricultural landscapes: spatial patterns Biosafety Res 2, 213-217. of contamination. Ecol Econ 53, 387– 401. [15] Witten, I.H., Frank, E. 2011. Data Mining: Practical [3] Debeljak M., Demšar D., Džeroski S., Schiemann J., Machine Learning Tools and Techniques - 3rd edition. Wilhelm R., Meier-Bethke S. 2005. Modeling outcrossing Morgan Kaufmann. 54 Knowledge Discovery from Complex Ecological Data: Exploring Syrphidae Species in Agricultural Landscapes Marko Debeljak Vladimir Kuzmanovski Veronique Tosser Aneta Trajanov Jožef Stefan Institute and Jožef Stefan Institute and ARVALIS-Institut du vegetal, Jožef Stefan Institute and Jožef Stefan International Jožef Stefan International 91 720 Boigneville, Jožef Stefan International Postgraduate School Postgraduate School France Postgraduate School Jamova 39, Ljubljana, Jamova 39, Ljubljana, +33 1 64 99 23 15 Jamova 39, Ljubljana, Slovenia Slovenia v.tosser@arvalis.fr Slovenia +386 1 477 3124 +386 1 477 3143 +386 1 477 3662 marko.debeljak@ijs.si vladimir.kuzmanovski@ aneta.trajanov@ijs.si ijs.si ABSTRACT exploring the biodiversity of the auxiliary species of the family Modelling structures and processes in ecosystems, such as Syrphidae (hoverflies), which are the major predators of Aphid biodiversity has always been a complex task. Most often, pest species. The enhancement of biological pest control in ecological empirical data are incomplete, inconsistent, agricultural fields helps reduce the use of pesticides and has an imbalanced or very complex and a lot of effort should be put overall positive impact on crop production and quality of the into preprocessing of such data in order to carry out meaningful environment. analyses and modelling. Our goal is to analyse and model the taxonomic and functional In this study, we are dealing with biological pest control in diversity of syrphid species in order to estimate their potential agricultural landscapes. The improvement of the natural performance of pest control in the studied area. regulation of organisms detrimental to agricultural production Quantifying and assessing the biological diversity of empirical through biological pest control has the potential to reduce the data collected in the fields is a non-trivial task and involves use of pesticides and has a positive impact on crop production calculations of different diversity measures, such as Hill and environment. Our research focuses on auxiliary species of numbers, Shannon’s and Simpson’s indexes, as well as quality the family Syrphidae, which control Aphid pest species. In checking and transformations of the available data in order to particular, our goal is to describe taxonomical and functional address the problem of modelling the taxonomic and functional diversity of syrphid species to assess the potential performance diversity of species. In addition, the available empirical data did of biological pest control in the studied area. not contain sufficient information needed for these calculations In this paper, we present the extensive process of data and transformations. Therefore, we needed to extract additional preprocessing for the purpose of modelling the taxonomic and data from several different datasets in order to cover all needed functional aspects of syrphid species. aspects of biodiversity. These calculations, data transformation and collation, are a Keywords necessary step towards preparing a high quality dataset, which Knowledge-discovery, data preprocessing, Syrphidae species, will enable us to obtain meaningful results and models. In this taxonomic and functional diversity, landscape structure, data paper, we present the complex data preprocessing task that was mining. carried out in order to explore and describe the taxonomic and 1. INTRODUCTION functional diversity of syrphid species. Knowledge discovery from ecological data is becoming an increasingly complex task. One reason for this is that ecosystems are very complex by themselves, representing 2. MATERIALS AND METHODS networks of interactions and interdependencies among its elements and the environment, which are most often difficult to 2.1 Data describe, explain or measure. The empirical data were provided by ARVALIS, Institut du vegetal. Data come from Boigneville and were collected in Empirical ecological data add another dimension to this 2009, 2010 and 2011. Samples of syrphid species were collected complexity. Namely, we are often faced with empirical with five Malaise and eight cornet traps (Figure 1) on a weekly ecological data that are incomplete, inconsistent, containing out- basis between March and November. Samples of caught insects of-range values, collected at different temporal and spatial have been determined to the species level and number of caught scales, dispersed in different databases, noisy and imbalanced specimens per species was counted. [3]. Therefore, special attention should be paid on the preprocessing and quality check of ecological data in order to To describe the ecological functional traits of syrphid species in obtain meaningful results, which makes the data preprocessing a Boigneville, we obtained an additional and extensive database very central step in the knowledge discovery process [7]. “Syrph The Net” [6]. It includes coded information on species’ macrohabitats, microsites, traits, range and status. The database In this study, we are dealing with the assessment of potential is updated annually it is used to analyse recorded species performance of biological pest control in agricultural fields in assemblages in relation to their habitat associations. the Boigneville area (Central France). In particular, we are 55 from 1.1.2008 to 31.12.2011, data about maximum, minimum, average temperature and cumulative rainfall have been collected at daily bases. 2.3 Data preprocessing Taxonomic and functional diversity of caught syrphid species was described by the total number of species, referred to as species richness and evenness (indicating how abundance is distributed among the species). Indices that combine species richness and evenness into a single value are referred to as diversity indices. a) Among the very large number of diversity indices, we decided to calculate the Hill diversity numbers (N0, N1, N2) because they are the easiest to interpret ecologically [4]. The three Hill numbers coincide with the three most important and known measures of diversity: S-number of species, H’-Shannon’s index and λ-Simpson’s index [3]. Shannon’s index H’ is calculated as: S  n n i   i  (S: number of species in the sample, H'  ln  1 n n i     ni: number of individuals of i-th species in the sample, n: number of all individuals in the sample). H’ is 0 if there is only one species and its value b) increases then both the number of species and their evenness increase. For a given number of species, the value of a Shannon Figure 1: Malaise (a) and cornet (b) traps for sampling diversity index is maximized when all types are equally syrphid species abundant. Data about landscape structure and crop properties were Simpson’s index is calculated as: 𝜆 = ∑𝑆 𝑝2 𝑖=1 𝑖 (pi: proportional described from the original dataset for a 500 m and 1000 m abundance of the i-th species pi= ni/N). Simpson’s index varies radius around the traps (Figure 2). In delineated area, the surface from 0 to 1 and if the community consists of only one species, of crops, natural vegetation (forests) and the length of linear Simpson’s index is 1 and there is no diversity. corridors (tree lines, grass strips, grass pathways, hedges, roads) The Hill numbers are given in units that represent the effective have been measured using GIS maps. Crop development stages number, i.e., the number of species in a sample, where each were estimated for each crop in the studied area. Landscape species is weighted by its abundance (N0=>N1=>N2): description includes absolute and relative coverage of the surrounding soil with crops. The coverage is expressed through - N0 is the number of all species in a sample i: N0 = n i; the percentage of groups of crops that share similar characteristics in response to Syrphids. - N1 is the number of abundant species and calculated from the Shannon’s diversity index H’: N1=eH'; - N2 is the number of very abundant species and is based on Simpson’s index λ: N2=1⁄λ. The distribution of abundance among the species is estimated by the evenness index. It is a ratio of observed to maximum diversity and it riches the highest values when the individuals are evenly distributed among species and it is independent of the number of species in the sample. Evenness is calculated as modified Hill number E5: 1 ( ) − 1 𝑁2 − 1 𝐸5 = 𝜆 = 𝑒𝐻′ − 1 𝑁1 − 1 In addition, we calculated several measures that describe the landscape diversity. Among the most popular metrics used to quantify the landscape composition are the Shannon’s index, which emphasizes the richness component of diversity, and Simpson’s index, which emphasizes the evenness component [5]. The Shannon’s index is therefore recommended for Figure 2: Area with 1000 m and 500 m radius around a landscape management within an ecological framework. sampling point for landscape and crop characterisation Simpson’s index is more responsive to the dominant cover type and is used for specific situations where one cover type is We obtained climatic data from the French national prevailing. These diversity indices can be used to evaluate: i) meteorological station located in Boigneville. For the period Landscape richness, which is simply the number of land cover 56 types present within the landscape; ii) Landscape diversity, (2009-2011). Abundance is a strong indicator about the syrphid which evaluates both richness and evenness aspects of the species, but it does not enable investigation of the correlation landscape; iii) Landscape evenness, which normalizes for the between taxonomic and functional diversity and landscape effect of richness on the diversity index. elements and crops. Thus, as with the biodiversity metrics, Hill numbers and The landscape structure of the studied area (radius of 500 and Evenness index were used, where the number of species was 1000 m) is very diverse and its diversity does not change over replaced with the number of crops, while the number of the studied period very much. Table 2 shows landscape diversity individuals of each species with the land cover area (m2). and evenness indices for a radius of 500 m around the traps. For each field the following soil properties have been described: soil texture class and content of clay, sand, coarse fragments, available water holding capacity, and bulk density. The data were acquired from dedicated soil database, established and maintained by ARVALIS. Phenological development of syrphids (e.g., egg, larval, pupal, and total development times), crops (e.g., leaf unfolding, grass growing, flowering of plants, etc.) and natural vegetation depend on temperature conditions, which are described by degree-days. Degree-days were calculated using simple logistic equation [1]: 12 𝛽 2 ∗ [𝛽 1ln⁡(𝑒𝛽3𝑡+𝛽2 + 1) 1𝑡 − ] 𝛽3 𝐷𝐷 = 0 − 𝑀𝑇𝑇 24 According to the study of [2], the minimum daily average temperature of syrphid species is 60C. For reliability reasons, we selected 50C for the minimum temperature threshold in our study. Introduction of degree days enables objective comparisons of abundance and diversity dynamic between years and locations, because climatic and environmental conditions for the same calendar date vary between years, therefore calendar days cannot be used as a temporal reference point. 3. RESULTS With regards to biological control, the abundance of predatory Figure 3: Relative rank abundance of syrphid species for individuals (predators and parasitoids) might be far more years 2009-2011 at three-week time steps (e.g., 09|w24-26 relevant for performing pest control than their diversity. In stands for week 24 to week 26 in year 2009) particular, the results of the abundance of the syrphid species in Table 2: Diversity of crops in the area within 500 m of the the studied agricultural landscape of Boigneville, show that four traps syrphid species significantly dominated over other caught species. Comparisons between years show mostly no differences Diversity indices Year of the top most abundant species nor their relative abundance at a yearly level (Table 1). Number of crops (NO) 2009 15.3 2010 14.0 Table 1: Relative rank abundance of syrphid species on 2011 13.2 annual level (2009-2011) Number of dominant crops (N1) 2009 7.5 Syrphid species 2009 2010 2011 2009-2011 2010 8.4 (%) (%) (%) (%) Sphaerophoria scripta 43.2 72.0 53.7 55.0 2011 7.2 Episyrphus balteatus 22.6 3.6 4.4 10.5 Number of very abundant crops (N2) 2009 5.5 Melanostoma mellinum 9.0 11.1 16.3 12.4 2010 6.5 Eupeodes corollae 8.2 3.6 4.4 10.4 2011 5.5 Other species 17.0 9.7 8.5 11.8 Evenness (E5) 2009 0.7 2010 0.7 The prevailing dominance of these four syrphid species has been 2011 0.7 confirmed also at three week time period (Figure 3). This indicates very low variability of syrhid species at both inter and The prevailing number of crops is high and the value of intra annual level, which further indicates stability of their living evenness is relatively high as well. This indicates an even conditions, which means that landscape structure and applied distribution of crops in the area. However, the prevailing crops crop management have not changed much in the studied period are cereals, where winter wheat covers the largest areas. 57 In addition to habitat variability, taxomomic (Table 3) and quality of their predictions. In our study, we used all standard functional (Table 4) diversity of syrphid species appear to be steps to ensure high quality of data, such as data cleaning, high and relatively stable at the inter-annual level. Hill number 2 outlier detection, missing value treatment, etc. (N2) for taxonomic diversity shows that only the three species Table 5: Groups and number of attributes in the final are very abundant, therefore the evenness values are rather low. dataset Table 3: Taxonomic diversity of syrphid species Group of attributes Number of attributes Year Abundance N0 N1 N2 Evenness Field description 13 2009 4844 57 4.89 3.18 0.560 Species description 7 2010 3748 48 3.68 2.61 0.599 Soil description 7 Landscape description 48 2011 5469 56 4.26 2.97 0.604 Meteorological conditions 7 Total 14061 84 4.56 2.98 0.558 Temporal component 4 Taxonomical aspect of species 84 Functional aspect of species 39 Table 4: Functional diversity of syrphid species Functional groups N0 N1 N2 Evenness The majority of our work was focused on transformation and Larvae: terrestrial 31 4.02 2.68 0.514 creation of new attributes in order to facilitate the knowledge discovery process about the potential contribution of syrphid Larvae: herbal layer 19 3.53 2.49 0.552 species to biological control of Aphid pest species. In order to Larvae: root zone 14 2.88 2.01 0.493 do this, we used extensive amounts of ecological knowledge about the description of taxonomic and functional properties of Overwinter 32 4.04 2.68 0.512 the study group of auxiliary species. The completeness and hibernation (OH) quality of the obtained (preprocessed) data were reviewed and OH: above ground 9 2.61 1.97 0.610 confirmed by ecological experts. surface OH: ground surface 22 3.12 2.25 0.555 We conclude that landscape and crop diversity support high taxonomic and functional diversity of syrphids. This is a very OH: root zone 11 5.68 3.90 0.575 promising preliminary approximation, which indicates that we Larval food: living can expect to obtain interesting results from data mining models. 6 1.98 1.60 0.558 plants Therefore, the next step is to apply various data mining methods Larval food: living on the preprocessed dataset in order to discover new knowledge 25 3.72 2.60 0.542 animals about interactions between the environment (landscape Adult food: nectar structure, crop management, soil, climate) and taxonomic and 32 4.04 2.68 0.513 flowers functional diversity of syrphid species. The new knowledge will be used for enhancing the existing syrphid species to perform Adult food: pollen 12 4.04 2.68 0.573 efficient biological pest control on growing crops. flowers The analysis of functional diversity of syrphid species shows 5. REFERENCES that larvae of 62% of caught species live in herbal layers and [1] Caicedo, D.R., Torres, J.M.C., Cure, J.R. 2012. larvae of 81% of species feed on living animals. This indicates Comparison of eight degree-days estimation methods in very high potential of syrphid species to perform biological pest four agroecological regions in Colombia. Agrometeorology control because most of syrphid species are feeding on living 71, 299-307. Aphids (aphidophagous), which are the major pest of cereal [2] Hassall, C., Owen, J., Gilbert, F. 2017. Phenological shifts crops. The majority (68%) hibernate on the ground, while all in hoverflies (Diptera: Syrphidae): linking measurement adults feed on both nectar and pollen. Such obligatory and mechanism. Ecography 40, 853–863. dependence on nectar and pollen food indicates that the [3] Legendre P., Legendre L. 2012. Numerical ecology. landscape and crop structure provides these required food Elsevier, Amsterdam, Netherlands. sources. [4] Ludwig, J.A., Reynolds, J.F. 1988. Statistical Ecology. Finally, after all the preprocessing of taxonomic, functional and New York, Chichester, Brisbane, Toronto, Singapore, John environmental data, we got a dataset comprising seven groups of Wiley & Sons. attributes describing: properties of the fields with sampling [5] Querner, P., Bruckner, A., Drapela, T., Moser, D., Zaller, traps, taxonomic and functional descriptions of caught syrphid J.G., Frank, T. 2013. Landscape and site effects on species, soil properties, descriptions of the landscape and crop Collembola diversity and abundance in winter oilseed rape properties, meteorological conditions with degree-days and fields in eastern Austria. Agr Ecosyst Environ 164, 145– descriptions of temporal components of the data collected. The 154. final set of attributes contains in total 209 attributes (Table 5). [6] Speight, M.C.D., Castella, E. 2016. StN Content and Glossary of terms. Syrph the Net, the database of European 4. CONCLUSIONS Syrphidae (Diptera), Vol. 94, 89 pp, Syrph the Net publications, Dublin. Preprocessing of data is a very important step in ecological [7] Witten, I.H., Frank, E. 2011. Data Mining: Practical modelling in general and data mining in particular, because the Machine Learning Tools and Techniques - 3rd edition. quality of input data affects the structure of the models and the Morgan Kaufmann. 58 A Comparison of DEXi, AHP and ANP Decision Models for Evaluation of Tourist Farms Tanja Dergan Institut Jožef Stefan Jamova cesta 39, 1000 Ljubljana Slovenia tanja.dergan@ijs.si ABSTRACT also to help potential consumers to make decisions about which tourist farm to choose. This research covers a comparison between decision models created with the DEXi tool based on the DEX methodology, and 2. DATA DESCRIPTION the decision models made by Super Decision tool using Analytical The study evaluates four tourist farms, whose locations are in hierarchical process (AHP) and Analytical network process (ANP) different geographical regions of Slovenia: I - Izola, II - Pohorje, methodology, for analysing decisions, on a case study of tourist III - Ponikva, IV - Logarska dolina. In addition to the location of farms. Based on the performance of decision models, the most the selected tourist farms, it was also important that they provide appropriate decision making methodology that maximally the possibility of overnight stays. At this stage, 6 criteria and 17 satisfied the evaluation criteria was DEX. Based on the empirical sub-criteria were selected, (Table 1), and used in all decision data, the best evaluated farm was Tourist farm IV from Logarska models. The criteria were chosen personally bay the researcher, dolina, which achieved the best evaluation by all three decision while the data for the evaluation of tourist farms were obtained models. through personal interviews of farm owners and guests, as well as Keywords: survey questionnaires of potential guests. The data obtained through the interviews of the farm owners and guests have given Decision support modelling, DEXi, Analytical hierarchical numerical values, while the survey questionnaires of potential process, Analytical network process, touristic farm Form of Offer: 1. INTRODUCTION Criteria Location supplementary complementary Decision making can be defined as a cognitive mind process, a activity activity human quality used to solve everyday situations. Some decisions -Where is it -Stationary -Food may be felt as easy, others difficult and demanding. The located farm -Drink development of the decision support system represents a new step Sub- -Accessibility -Excursion -Sport activities towards optimisation and improvement of the whole decision criteria farm -Living space making process [1]. -Tourist farm Multi-criteria decision making is an approach where we make Logo decisions on the basis of several criteria. This approach is The Host to guest necessary when intuitive decision making is not sufficient, either Criteria surrounding of Hospitality relationship because of conflicts between criteria, or because of the differences the farm between decision makers. -Flowers and -Reception -Family greenery -Events arrangement Tourist farms provide an important development potential for the Sub- -Preserving the -Access to -Cleanliness inhabitants of the Slovenian countryside [2]. There are currently criteria cultural information over 800 tourist farms in Slovenia and their number is growing landscape for guests rapidly. The increasing number of tourist farms brings many -Categorization benefits for the regional development and local inhabitants such guest resulted in descriptive data. as prevention of emigration of young people from the countryside, preservation of the cultural landscape and the provision of social Table 1: Structure of data. security for farming families with an additional source of income. The problem that we deal with, relates to the selection of an 3. DECISION MODELS appropriate modelling method for the case study of tourist farms Different methods such as Servqual and Dematal [5] can be used and the criteria that they should fulfil in order to be chosen by to measure the quality of tourist services. Some of them also users (tourists), for spending their holidays. enable work with inaccurate and incomplete data and use an interval account such as Mund, Promethee [5]. However, in this The goal of this article is to assess the decision models built with study we focused on the decision methods for construction of three different decision making methodologies, which is based on DEXi, AHP and ANP decision models. comparisons of their complexity, interdependency and consistency. The objective of these decision making models was 59 3.1 DEXi elements with each other. The results in the AHP method are interpreted in three ways: i) ‘Normals’, where the results are DEX is a multi-attribute methodology for decision making. The methodology is based on attributes with a finite set of qualitative presented in the form of priorities, where each one of the values instead of attributes with numerical values [8]. The DEX alternatives are summed and then each element is divided by the sum, ii) ‘Ideals’, where the values are obtained from ‘Normals’ by methodology enables a construction of transparent and comprehensive models and it provides techniques for integration dividing each of its entries by the largest value in the column, so of attributes through aggregation rules in form of hierarchical that the best alternative gets a priority of 1 and the others get proportion less than 1, and iii) ‘Raw’, whose values are read decision trees. directly from the Limit Supermatrix. DEXi is a software modelling tool, which is based on the DEX methodology and facilitates the development of qualitative Multi 3.3 Analytical Network Process (ANP) Attribute Decision Models (MADM) and enables an evaluation The ANP is implemented in the software Super Decisions and has and what-if analysis of decision options [8]. DEXi is useful in been applied to various problems both to deal with decisions and cases where we do not have numerical data or ratings, but only to illustrate the uses of the new theory. The ANP is a coupling of qualitative ones [3]. In general, DEXi models are customised and two parts. The first consists of a control hierarchy or network of do not have a complex structure, are insensitive to minor changes criteria and subcriteria that control the interactions in the system in input data and capable of resetting procedures [6]. under study. The second is a network of influences among the In the DEXi modelling tool, the alternatives are described by elements and clusters [7]. In the Super Decision modelling tool, initial attributes, which are then evaluated separately according to the criteria were grouped into a network model, with clusters and their values. The final evaluation of the alternatives is obtained by with related criteria, and not in a hierarchical level. The method an aggregation process of input data (values of initial attributes allows for interactions and feedbacks within the cluster, as well as Xi) using aggregation functions Fi. The output value of the between clusters, for example the alternatives of the decision in topmost node in the decision tree (decision model) is used for another cluster. This helps to make the basic computer operations selection of the most suitable alternative among all evaluated and logical multiplication in different ranges as required by the alternatives. model. The mutual pairwise comparisons were performed like in AHP model, base on the Saaty scale [7] from 1 to 9 meaning: 1) The DEXi model was applied to evaluate four tourist farms using criteria are the same, 2) criteria is equivalent to another or has a data derived from interviews, as well as survey questionnaires. moderate advantage over another, 3) criteria has a moderate The tourist farms and their regulatory standards were precisely advantage over other criteria, 4) criteria has a moderate to great defined, in order to select attributes, which have been structured advantage over another criteria, 5) criteria has a great advantage in the DEXi model. The goal of the model was to decompose the over another criteria, 6) criteria has a great advantage over other problem into smaller sub-problems, which were assessed criteria, 7) criteria has a very great advantage over another individually using criteria determined by the decision maker: For criteria, 8) criteria has a very large to an extremely high advantage example, the set of values for the attribute "Sport Activities" was: over another criteria and 9) criteria has an extremely high excellent; medium and poor. Through the process of hierarchical advantage over another criteria. Our research focuses on integration, using the utility functions obtained provided by the development of three ANP model applications: i) the simplest decision maker, the final assessment of the top-most attribute was single model, which was built only in one layer and was the determined. The outcome of the evaluation was an assessment of easiest to build, ii) the two-layers model, which divide the model tourist farms. into upper and lower levels, and iii) the complex three-layers 3.2 Analytical hierarchical process (AHP) model, which consists of several layers of sub-networks and is one of the most demanding models in the presented research. The AHP is an established and well-researched method of analysing a input data, as well as the data from interviews and surveys used in hierarchical decision-making processes based on mathematics and the ANP model was excerpt from the previously presented AHP psychology [7]. The model was built in the Super Decision and DEXi models. However, due to the complexity of the ANP modelling tool [4] and consists of a general goal (selection of the complex three-layers model, the criteria are further subdivided to best tourist farm), criteria and sub-criteria (Table1) and common create a more extensive model. options or alternatives (Tourist farm I, Tourist farm II, Tourist farm III and Tourist farm IV). The structure enables the 4. RESULTS WITH DISCUSSION possibility for taking into account the given elements at the selected level as well as all elements at lower levels. The criteria The aim of the research was to compare DEXi, AHP and ANP and their hierarchical structure are the same in the DEXi models multi-criteria models, for an evaluation of tourist farms. as well as in AHP models, which provides the basis for 4.1 DEXi model comparisons of the models. The tourist farm is an alternative and In DEXi modelling, the main focus is on the rationality and the therefore lies the highest in the hierarchy tree. They are regularity of the criteria. Based on the obtained data, a multi- determined by subordinate criteria, and further by lower sub- criteria model was developed that maximally met the given criteria. The method converts the evaluation of tourist farms into criteria. The best evaluated tourist farm (Figure 1) is from numerical values that can be processed and compared for each Logarska dolina (Tourist farm IV). The second best tourist farm criteria in the hierarchy. Mutual pairwise comparisons of was Tourist farm II, which achieved the same level of evaluation alternatives (tourist farms) were performed in a hierarchical model scores in almost all criteria as Tourist farm IV. Tourist farm II was based on the obtained data (surveys, interviews). A basic scale evaluated worse only in the criterion “Complementary activities” (i.e., the Saaty scale) from 1 to 9 was used, where each gives a (Figure 1). Tourist farm IV and Tourist farm II got the highest specific preference [7]. The use of numerical weights allows for score in the criteria “The surrounding of the farm”. However, in rational and consistent comparison of different or incompatible the criteria “Tourist Farm Logo” and “Where is it located” they 60 did not receive the best estimates. Tourist farm III was evaluated for the criteria “The surrounding of the farm”. The Tourist farm I well and it is potentially a good choice for tourists. The worst was inadequate due to poor estimates of the criteria evaluated was Tourist farm I, although it was very well evaluated “Complementary activities” and “Hospitality” (Figure 2). Figure 1: Evaluated criteria in DEXi model for Tourist farm II and Tourist farm IV. Figure 2: Evaluated criteria in DEXi model for Tourist farm I and Tourist farm III. 4.2 AHP model The criteria “Complementary activity” and “Location” contributed In AHP method, the development of decision-making model the highest values, while the least impact on the final results had (identification of the decision problem, identification of the criteria “Form of supplementary activity”. alternatives and determination of criteria) was similar as in DEXi. The difference in AHP is in the pairwise comparison of the 4.3 ANP model criteria with respect to the goal and the pairwise comparison of Three different models have been developed with the ANP the alternatives. The results of the AHP model (Figure 3), based method in the Super Decision tool. According to the input data on the Normals values show that the best evaluated tourist farm and the problems that were considered, we came to the conclusion was from Logarska dolina (Tourist farm IV) which received 40%. (Table 2) that the outputs for all three applications of the ANP Tourist farm II received 37%, Tourist farm I received 13 % and methods show that the Tourist farm IV from Logarska dolina was Tourist farm III received the lowest percentages 10%. Based on selected as the most appropriate one. The tourist farm IV received Ideals values (Figure 3), the results can be interpreted also in the the highest percentage in the three layered model (40%), in the way as: Tourist farm I is 33% as good as a Tourist farm IV, two layered model 38% and in single layered model 36%. The Tourist farm II is 92% as good as Tourist farm IV and Tourist lowest percentage achieved Tourist farm III in the three layered farm III is 25% as good as Tourist farm IV. model (10%) and 17% in both two layered and single layered model. The results in the simplest single layered model, showed that the greatest impact on the final results had criteria Tourist farm IV “Complementary activity” and “The surrounding of the farm”, Tourist farm III while the criteria “Location” heed the least influence on the Tourist farm II results. In the two-layereds model, Tourist farm IV got the best assessment in all criteria. Tourist farm II received good Tourist farm I assessment of the criteria “Host to guest relationship”, “Form of supplementary activity”, “Hospitality” and “Complementary 0 0,5 1 1,5 RAW NORMALS IDEALS activity”. Tourist farm I received good assessment for the criteria “Location2 and “Form of supplementary activity”. Tourist farm Figure 3: Synthesized priorities for the alternatives in AHP III was assessed with lower estimates, only a slightly higher rating model. according to the criteria “The surrounding of the farm”. In the 61 complex three-layereds model, Tourist farm IV achieved the best numerical values. The ANP model build in the Super Decision assessment in the criteria “Costs” and “Priority”. Among all the tool proved to be more complicated. The application of the simple criteria, three were selected that were assessed as most important single network model and the two-layers model are and that had the greatest impact on the final results: “Where is it comprehensive, and still sufficiently understandable, while the located”, “Stationary farm” and “Cleanliness”. By comparing approach to build a complex three-layers model is much more results of all three ANP model applications, a deviation in the complex. Despite the fact that, due to its size, the three-layers percentages was observed. The percentages of the three layered model gives us more precise results, its use is more difficult. model were for Tourist farm IV and Tourist farm II evaluated This led to the conclusion that in order to create a model with higher and for Tourist farm III and Tourist farm I the percentages qualitative attributes, perform what-if analysis and include were evaluated lower in comparison with the percentages of the additional alternatives for evaluation (e.g., Tourist farm 5, Tourist single and two layered models, where only a smaller deviation of farm 6., etc.), the most appropriate model was built in DEXi. The the percentage occurred (Table 2). The three layered model AHP model, which is based on numerical values of input data, can contained in the upper level two new clusters, Strategic and Basic. be more precise then the DEXi model, but on the other hand, Strategic served in the model for evaluating the Basic criteria determining one value in the comparison matrix in AHP, is much using the rating model. The basic cluster consists of additional more difficult then determining it in DEXi models. Thus, if we criteria: “Priority”, “Cost” and “Risk”. The new criteria contain want to use quantitative attributes without the evaluation of lower levels, where the already known criteria and sub-criteria additional alternatives, the AHP methods could be applied. The used in DEXi and AHP models are added (Table1). The criteria application of the three-layers ANP model is non-transparent due were grouped into i) Priority criteria: “Location” and to multiple layers and its construction is highly time-consuming. “Hospitality”, ii) Cost criteria: “Complementary activity” and The use of a single or two-layer ANP model in the Super Decision “Form of supplementary activity” and iii) Risk criteria: “Host to toll can give sufficiently more precise results. guest relationship” and “The surrounding of the farm”. The presented expansion of the three-layered model led to a deviation Each of the assessed decision modelling methods has its of the percentage (Table 2). advantages and disadvantages. However, we have found that DEXi is the best modelling approach for the assessment of tourist Table 2: ANP total results. farms, and that the Tourist farm IV located in Logarska dolina represents the best provider, which was not confirmed only by the ANP-Single ANP- Two ANP-Three DEXi model, but also by the models developed in Super Decision layered model layered model layered model tool with AHP and ANP method. Tourist farm I 21% 20% 15% Tourist farm II 26% 25% 35% 6. REFERENCES Tourist farm III 17% 17% 10% [1] Babčec R. September 2010. Analysis functionality and Tourist farm IV 36% 38% 40% system performance to support group decision-making thesis. Faculty of organizational sciences, UM. + [2] Deklava M. 1998. Development of tourism places. 5. CONCLUSIONS Participation of the villagers. Ljubljana, Tourist Association The goal of this study was comparison of the DEXi tool in the DEX methodology, and the Super Decision tool in the AHP and [3] Dergan T. 2010. Development of Multi-criteria model for ANP methodology. The study presented three examples of tourism farm assessment. Graduate thesis. UM. + decision models for evaluation of tourist farms, which in addition [4] Dergan T. 2014. Development and application of AHP and to the final results, also provided appropriate measures to improve ANP decision models for farm tourism analysis. Master’s the offers of the more poorly assessed alternatives. The objective thesis. University Maribor of these decision making models used in the survey was also to [5] Jereb E, Bohanec M, Rajkovič V. 2003. DEXi computer help potential consumers to make decisions about the most program for multi-parameter decision making. Kranj, appropriate tourist farms. The results of all models indicated that Modern organization. the Tourist farm IV located in Logarska dolina received the best evaluation results and represents the best provider of touristic [6] Murn T. 2005. Learning for the Future Multi-Attribute activities in the farm. The applied methodologies were found to be Decision making Models and the sistem DEXi for schools. very successful and effective. DEXi tool has proved a very simple Karlstad, City Tyrick: 329-343 str. model, which is not related to numerical values of input data. It is [7] Saaty R.W. 2003. Decision making in complex environments also easy to add additional criteria to a structured decision tree, – Analytical hierarchical process (AHP) for Decision Making despite the measurements of input data and integration functions. and Analytical network process (ANP) for Decision Making The modelling software indicates where model needs to be with Dependence and Feedback. University of Pittsburgh. modified or corrected due to an additionally added criterion. AHP [8] Zadnik S.L, Žerovnik J, Kljajić B.M, Drobne S. 2015. in the Super Decision tool also presents a simple and transparent Proccedings Sor 15. The 13th International Symposium on way to build a model and represents a very good alternative to the operational research. Bled. Slovenia DEXi model. However, in contrast to DEXi, it is bound to 62 A State-Transition Decision Support Model for Medication Change of Parkinson's Disease Patients Biljana Mileva Boshkoska1,2 Dragana Miljković1 Anita Valmarska1,4 biljana.mileva@ijs.si dragana.miljkovic@ijs.si anita.valmarska@ijs.si Dimitris Gatsios3 George Rigas3 Spiros Konitsiotis3,5 dgatsios@cc.uoi.gr george.a.rigas@gmail.com skonitso@gmail.com Kostas M. Tsiouris3 Dimitrios Fotiadis3 Marko Bohanec1 tsiourisk@biomed.ntua.gr fotiadis@cc.uoi.gr marko.bohanec@ijs.si ABSTRACT duration), mental problems (impulsivity, cognition, hallucinations and paranoia), epidemiologic data (patient’s age) and In this paper, we present a state-transition decision support model for medication change of patients with Parkinson's disease (PD), comorbidities (cardiovascular problems, hypertension and low implemented with method DEX. Today, PD patients can be blood pressure). The model is composed of (1) a state-transition treated with three basic medications: levodopa, dopamine agonist, model that presents the medication change among levodopa, MAO-B inhibitors, and their combinations. We propose a model dopamine agonist, MAO-B inhibitors and their combinations, and which, based on the current patient’s symptoms (motor symptoms, (2) decision rules for triggering the changes, represented in terms mental problems, epidemiologic data and comorbidities), suggests of a qualitative multi-criteria model. The latter has been how to change the medication treatment given the patient’s developed using the DEX method [4], which integrates the current state. The model is based on expert’s knowledge of qualitative multi-criteria decision modeling with rule-based expert neurologists and is composed of (1) a state-transition model that systems. presents all possible medication changes, and (2) decision rules for triggering the changes, represented in terms of a qualitative 2. MODEL DESIGN rule-based multi-criteria model. The model assesses all states The model development was performed with neurologists who described by the state-transition model and proposes multiple work with PD patients. The process of decision analysis led to a different yet still correct possibilities for medication change. design of a model composed of two key elements: (1) A state- transition model that represents all possible combinations of used Categories and Subject Descriptors medicaments and transitions between them, and (2) a multi- H.4.2 [Types of Systems]: Decision support. criteria DEX model that provides decision rules for each J.3 [Life and Medical Sciences]: Medical information systems. transition. General Terms 2.1 A state-transition model Algorithms, Management, Measurement, Design, Human Factors In the state-transition model the medication treatments (states) and transitions among them are represented in a form of a cube as Keywords presented in Figure 1. In Figure 1, each medication-treatment state Parkinson’s disease, medication change, decision model is as a circle and each change of medication treatment is represented with a directed arc. Each state corresponds to the set 1. INTRODUCTION of medications that constitute the current treatment. The set can be Parkinson's disease (PD) is a complicated, individual degenerative empty (the symbol O indicates no medication therapy), or can disorder of the central nervous system for which there is no cure. consist of any combination of DA, LD and E (Enzymes, such as Hence it requires a long-term, interdisciplinary disease MAO-B inhibitor). For example, the state DA+E means that the management including typical medicament treatment with current medication treatment of the patient consists of dopamine levodopa (LD), dopamine agonist (DA), and enzymes (E), such as agonist (DA) and MAO-B inhibitor. From this state there are three possible state changes depending on the combinations of patient’s MAO-B inhibitor. Due to the different combinations of motor and mental symptoms from which PD patients suffer, in addition to symptoms: add LD to the treatment (state denoted as LD+DA+E), existing comorbidities, the interchange of medications and their remove DA from the current treatment (state E) or remove E and combinations is patient-specific [1]. In the framework of the EU use only DA (state DA). Horizon 2020 project PD_manager (http://www.parkinson- The absence of a directed arc between two states means that a manager.eu/) [2] we developed a decision support model, called particular change of medication treatment is not addressed in the the “How” model, for PD management which suggests how to model, either because it has been deliberately excluded change the medication treatment given patients’ current state. The (transitions from and to state O, which are out of scope of the assessment is based on data that include patients’ motor PD_manager project), or is rarely or not at all used in practice. A symptoms (dyskinesia intensity, dyskinesia duration, offs 1Jožef Stefan Institute, Ljubljana, Slovenia 2Faculty of information studies in Novo mesto, Slovenia 3Unit of Medical Technology and Intelligent Information Systems, University of Ioannina, Greece 4Jožef Stefan International Postgraduate School, Ljubljana, Slovenia 5Department of Neurology, Medical School, University of Ioannina, Ioannina, Greece 63 reflexive arc means an increase/decrease of the medication Aggregation of the basic attributes leads to two sets of attributes. (dosage or intake) [3]. The first set is composed of six aggregated attributes: Dyskinesia, MotorSymptoms, CurrentTherapy, PersonalCharacteristics, Comorbidities and MentalProblems. The purpose of this set of attributes is to aggregate several specific symptoms into common indicators, which are used as inputs to the second set of aggregated attributes. For instance, Dyskinesia is a common indicator of patient’s involuntary movements caused as a side effect of medications; it is determined by aggregating the basic attributes offs duration, d yskinesia intensity, and dyskinesia duration. The second group of aggregated attributes forms a set of 15 submodels, which determine the transitions from one medication state to the other one as given in the state-transition diagram (Figure 1). Those submodels are the following: Figure 1: A state-transition model for medication change among levodopa (L), dopamine agonist (DA), MAO-B 1. ChangeDAtoLD: Change therapy from dopamine agonist to inhibitors (E) and their combinations. Symbol O represents levodopa the state where the patient does not take medications. 2. ChangeDAtoDA+LD: Change therapy from dopamine agonist to dopamine agonist and levodopa 3. ChangeDAtoDA+MAOI: Change therapy from dopamine 2.2 DEX model agonist to dopamine agonist and MAO-B inhibitors 4. DecreaseDAdosage: Decrease the dosage of dopamine The transitions in the state-transition model (Figure 1) are agonist triggered according to a multi-attribute model, which is 5. IncreaseDAdosage: Increase the dosage of dopamine responsible for interpreting patients’ motor symptoms, mental agonist problems, epidemiologic data and comorbidities, and aggregating them into an overall assessment of the potential medication 6. ChangeLDtoLD+DA: Change therapy from levodopa to changes of a given patient. The model is hierarchical and levodopa and dopamine agonist qualitative, developed using a qualitative multi-attribute 7. IncreaseLDdosage: Increase the dosage of levodopa modelling method DEX [4]. DEX models decompose the decision 8. IncreaseLDintake: Increase the intake of levodopa problem into smaller, less complex sub problems, which are 9. DecreaseLDintake: Decrease the intake of levodopa represented by a hierarchy of attributes. Attributes from the 10. DecreaseLDdosage: Decrease the dosage of levodopa decision alternatives are aggregated in order to obtain an overall 11. ChangeDA+LDtoLD: Change therapy from dopamine the evaluation or recommendation. DEX belongs to the class of agonist and levodopa to levodopa qualitative multi-criteria decision making methods: it uses 12. ChangeMAOItoMAOI+DA: Change therapy from MAO-B qualitative (discrete) variables instead of quantitative (numerical) inhibitors to MAO-B inhibitors and dopamine agonist ones, and employs decision rules rather than numerical 13. ChangeMAOItoMAOI+LD: Change therapy from MAO-B aggregation functions for the aggregation of attributes. The inhibitors to MAO-B inhibitors and levodopa method DEX is supported by DEXi [5], freely available software 14. StopMAOI: Stop using MAO-B inhibitors that supports both the development of DEX models and their 15. AddMAOI: Add MAO-B inhibitors to the current therapy. application for the evaluation and analysis of decision alternatives. DEX was chosen for modelling due to its previous At the top of each submodel, there is the root attribute which successful usage for implementation of decision support models in represents the overall assessment of medication change under health care [6][7]. consideration. For example, the submodel ChangeDA+LDtoDA Using DEX principles of model development, the state-transtion estimates the change of medication from dopamine agonist and model from Figure 1 is mapped into a qualitative multi-attribute levodopa to dopamine agonist based on the information whether model presented in Figure 2. The model consists of basic and the patient already takes DA ( usingDA) and LD ( usingLD), if the aggregated attributes given in a structure that identifies possible patient has increased mental problems ( MentalProblems) and/or transitions in the state-transition diagram for a given patient [8]. cardiovascular problems ( cardiovascular). The model combines 22 basic attributes including data about All submodels were obtained through expert modelling. In this motor symptoms (rigidity, tremor, and bradykinesia), mental case, decision-support models were developed in collaboration problems (impulsivity, cognition, hallucinations, paranoia), between the neurologists (experts) from and the decision analyst. comorbidities (cardiovascular, low blood pressure, hypertension), The work proceeds in the form of a question-answer dialogue, led and dyskinesia (offs duration, intensity, and duration). In addition, by the analyst, aimed at identifying the important indicators and there is data about patient’s age and activity, and data about the decision rules used, implicitly or explicitly, by the expert when current therapy (which medications is the patient currently using, making decisions. and whether or not the maximum dosages of DA and LD have been reached). The values of these attributes constitute model’s inputs. 64 DEXi MedicationChangeStates6.dxi 14.3.17 Page 1 DEXi MedicationChangeStates6.dxi 14.3.17 Page 2 Scales Attribute Scale ChangeDA+LDtoLD yes; no ChangeDAtoLD yes; no usingDA yes; no usingDA yes; no usingLD yes; no MentalProblems yes; no MentalProblems yes; no cardiovascular yes; no cardiovascular yes; no low blood pressure yes; no ChangeMAOItoMAOI+DA yes; no ChangeDAtoDA+LD yes; no usingMAOI yes; no usingDA yes; no MotorSymptoms yes; no maxDA yes; no MentalProblems yes; no MotorSymptoms yes; no ChangeMAOItoMAOI+LD yes; no ChangeDAtoDA+MAOI yes; no usingMAOI yes; no usingDA yes; no MotorSymptoms yes; no MotorSymptoms yes; no StopMAOI yes; no cardiovascular yes; no usingMAOI yes; no DecreaseDAdosage yes; no usingDA yes; no usingDA yes; no Dyskinesia yes; no MentalProblems yes; no MentalProblems yes; no cardiovascular yes; no hypertension yes; no low blood pressure yes; no AddMAOI yes; no IncreaseDAdosage yes; no usingMAOI yes; no usingDA yes; no usingDA yes; no maxDA yes; no usingLD MotorSymptoms yes; no yes; no offs duration yes; no offs duration yes; no MentalProblems yes; no MotorSymptoms yes; no cardiovascular yes; no MotorSymptoms yes; no age lt65; 65-75; gt75 rigidity yes; no activity yes; no Tremor yes; no ChangeLDtoLD+DA yes; no tremor at rest yes; no usingLD yes; no action tremor yes; no MotorSymptoms yes; no postural tremor yes; no offs duration yes; no bradykinesia yes; no MentalProblems yes; no MentalProblems yes; no age lt65; 65-75; gt75 impulsivity yes; no IncreaseLDdosage yes; no cognition yes; no usingLD yes; no Psychosis yes; no maxLD yes; no hallucinations yes; no MotorSymptoms yes; no paranoia yes; no MentalProblems yes; no Comorbidities yes; no dyskinesia duration yes; no cardiovascular yes; no dyskinesia intensity yes; no low blood pressure yes; no offs duration yes; no hypertension yes; no DecreaseLDdosage yes; no Dyskinesia yes; no usingLD yes; no offs duration yes; no MotorSymptoms yes; no dyskinesia intensity yes; no MentalProblems yes; no dyskinesia duration yes; no dyskinesia intensity yes; no PersonalCharacteristics inactive; active dyskinesia duration yes; no age lt65; 65-75; gt75 offs duration yes; no activity yes; no IncreaseLDintake yes; no CurrentTherapy max; yes; no usingLD yes; no usingMAOI yes; no maxLD yes; no usingDA yes; no MotorSymptoms yes; no usingLD yes; no MentalProblems yes; no maxDA yes; no dyskinesia intensity yes; no maxLD yes; no dyskinesia duration yes; no offs duration yes; no DecreaseLDintake yes; no usingLD yes; no MotorSymptoms yes; no MentalProblems yes; no dyskinesia intensity yes; no dyskinesia duration yes; no offs duration yes; no Figure 2: Structure and value scales of the “How” medication change model Figure 2 shows the value scales and structure of the model. It IncreaseLDdosage aggregate attribute depends on seven lower shows that most attributes in the model are binary, each taking level attributes that correspond to current patients’ medication one of the two corresponding values: yes or no. Coloured values treatment and symptoms. These attributes are binary, so there are indicate that the corresponding attribute is ordered from left-to- 27 = 128 possible combinations of their values. The right, so that the leftmost (red) value indicates a problematic, and DEXi software was used to represent, manage and define such the rightmost ( green) a non-problematic patient’s condition. The combinations in the form of decision tables. All decision rules red/left values generally indicate a problem that should be contained in the model are presented in a tabular form together addressed by medication change. with a verbal interpretation. Table 1 is an example of a decision table that defines the decision rules for the aggregated attribute 2.3 Decision rules ChangeDatoLD. The symbol ‘*’ used in the decision tables For each aggregate attribute in the DEX model, it is necessary to denotes any value that can appear at that position. For instance, in define the values of that attribute for all possible combinations of connection with an attribute than can take the values “yes” and lower-level (input) attribute values. For example, the “no”, the ‘*’ stands for “yes or no”. 65 According to the decision rules presented in Table 1, one may Research and Innovation Horizon 2020, under grant number read that the change of medication treatment from DA to LD 643706. should happen only when the patient already takes DA ( usingDA). The change may take place in in three different cases: the patient 5. REFERENCES has mental problems, cardiovascular problems, or low blood [1] Gatsios D, Rigas G, Miljković D, Koroušić-Seljak B, pressure. Otherwise, the change to LD should not happen. Bohanec M, Arredondo MS, Antonini A, Konitsiotis S, The whole model contains 21 other decision tables such as Table Fotiadis DI. 2016. Mhealth platform for Parkinson's disease 1, corresponding to the remaining aggregate attributes in the management. CBHI, 18th International Conference on model. Biomedicine and Health Informatics, February 25-26, 2016, Dubai, UAE. [2] PD_manager: m-Health platform for Parkinson’s disease Table 1: Decision rules for submodel ChangeDatoLD management. EU Framework Programme for Research and low Innovation Horizon 2020, Grant number 643706, 2015– Mental- cardio- Change- usingDA blood 2017, http://www.parkinson-manager.eu/ Problems vascular DAtoLD pressure [3] Bohanec, M, Antonini, A, Banks, A, Fotiadis, D.I, Gasparoli, E, Gatsios, E, Gentile, G, Hovhannisyan, T, 1 yes yes * * yes Konitsiotis, S, Koutsikos, K, Marcante, A, Mileva 2 yes * yes * yes Boshkoska, B, Miljković, D, Rigas, G, Tsiouris, K.M, Valmarska, A. Decision support models. PD_manager 3 yes * * yes yes project, Deliverable D5.2, 2017 4 * no no no no [4] Bohanec M, Rajkovič V, Bratko I, Zupan B, Žnidaršič M. 5 no * * * no DEX methodology: Three decades of qualitative multi- attribute modelling. Informatica 37, 2013, 49–54. [5] Bohanec, Marko . DEXi: Program for multi-attibute decision 3. CONCLUSIONS AND FUTURE WORK making user's manual : version 5.00, (IJS delovno poročilo, Using the DEX method, we developed a state-transition model 11897). 2015. and decision rules for medication change of PD patients. This [6] Bohanec M, Zupan B, Rajkovič V. 2000. Applications of approach assured that the model fulfils the following important qualitative multi-attribute decision models in health care, characteristics: completeness (it provides outputs for any possible International Journal of Medical Informatics; 58- 59:191- inputs), robustness (it works even if some input data is missing), 205 consistency (the model is free of logical errors), transparency (the model is fully “open” for the inspection of contained decision [7] Šušteršič, O, Rajkovič, U, Dinevski, D, Jereb, E, Rajkovič, rules), comprehensibility (the embedded decision rules are easy to V. 2009. Evaluating patients' health using a hierarchical understand and explain). The model assess all combinations of multi- attribute decision model. Journal of international possible medication changes that arise from the state-transition medical research; 37(5):1646-1654. model thus allowing interpretation of several different and yet [8] Mileva-Boshkoska, Biljana, Miljković, Dragana, Valmarska, correct scenarios for medication change for patients that suffer Anita, Gatsios, Dimitros, Rligas, George, Konitsiotis, from PD. Spyros, Tsiouris, Kostas M., Fotiadis, Dimitrios I., Bohanec, The future work in the framework of the PD_manager project will Marko. 2017. Decision support system for medication be focused on model evaluation and implementation. We intend to change of Parkinson's disease patients : a state-transition verify and validate the model on (1) real-life examples of model. V: LINDEN, Isabelle (ur.). Proceedings of the 2017 medication-change decisions, such as the Parkinson Progression International Conference on Decision Support System Marker Initiative dataset [9], (2) on real case patient’s scenarios, Technology, ICDSST 2017, with a theme on Data, (3) and in comparison with neurologists from different EU Information and Knowledge Visualisation in Decision countries. The model will be integrated in the PD_manager m- Making, 29-31 may 2017, Namur, Belgium. health platform for Parkinson’s disease management [2]. [9] Parkinson Progression Marker Initiative: The Parkinson Progression Marker Initiative (PPMI). Progress in 4. ACKNOWLEDGMENTS Neurobiology 95(4), 2011, 629–635. The work of the authors was supported by the PD_manager project, funded within the EU Framework Programme for 66 Designing a Personal Decision Support System for Congestive Heart Failure Management Marko Bohanec, Erik Dovgan, Paolo Emilio Puddu, Anneleen Baert, Pavel Maslov, Aljoša Vodopija, Michele Schiariti, Sofie Pardaens, Mitja Luštrek Maria Costanza Ciancarel i Els Clays Jožef Stefan Institute Sapienza University of Rome Ghent University Department of Intel igent Systems, Department of Cardiovascular, Department of Public Health Department of Knowledge Respiratory, Nephrologic and Geriatric De Pintelaan 185 – 4K3, 9000 Gent, Technologies Sciences Belgium Jamova cesta 39, 1000 Ljubljana, Piazzale Aldo Moro 5, Roma 00185, {anneleen.baert, Slovenia Italy sofie.pardaens, {marko.bohanec, erik.dovgan, {paoloemilio.puddu. els.clays}@ugent.be pavel.maslov, aljosa.vodopija, michele.schiariti}@uniroma1.it mitja.lustrek}@ijs.si mcostanza.ciancarel i@gmail.com ABSTRACT medical practices [3] and is designed so that it never suggests In this paper, we describe the design of the HeartMan Decision anything that would harm the patient. Support System (DSS). The DSS is aimed at helping patients In this paper, we present the design of the HeartMan Decision suffering from congestive heart failure to better manage their Support System (DSS), which was finalised in June 2017 [3]. In disease. The support includes regular measurements of patients’ section 2, we describe the overall functionality and architecture of physical and psychological state using a wristband and mobile the system, and define the roles of its modules that address (1) device, and providing advice about physical exercise, nutrition, physiological measurements, (2) physical exercise, (3) nutrition, medication therapy, and environment management. In the paper, (4) medication, (5) environment management, and (6) an overall architecture of the DSS is presented, followed by a management of calendars and plans. In section 3, we focus on the more detailed description of the module for physical exercise physical exercise module and present its most important management. components for (1) patients’ physical capacity assessment, (2) weekly exercise planning, and (3) daily exercise management. Categories and Subject Descriptors H.4.2 [Types of Systems]: Decision support. 2. DESIGN OF THE HEARTMAN DSS J.3 [Life and Medical Sciences]: Medical information systems. The HeartMan DSS aims at providing medical advice to CHF patients using predictive models, clinical care guidelines and General Terms expert knowledge. The purpose of a typical DSS is to passively Algorithms, Management, Measurement, Design, Human Factors present information to decision makers so that they can make maximally informed decisions. This DSS, however, is intended Keywords for patients who have limited medical knowledge and are consequently expected to follow guidelines with little discretion. Decision Support System, Personal Health System, Congestive Because of that, the DSS actively provides advice to patients, Heart Failure, Physical Exercise, Decision Models although it does offer choice where appropriate. In this way, it belongs to the category of cooperative DSS [4]. 1. INTRODUCTION Congestive heart failure (CHF) occurs when the heart is unable to pump sufficiently to maintain blood flow to meet the body's needs Health Phone Data Web [1]. Symptoms include shortness of breath, excessive tiredness, devices sensors management interface and leg swelling. CHF is a common, chronic, costly, and potentially fatal condition [2]. In 2015 it affected about 40 million people globally. In developed countries, around 2% of adults have Decision Support Modules Monitoring physical heart failure, increasing to 6–10% in age over 65. and psychological 1. Physiological state HeartMan (http://heartman-project.eu/) is a research project measurements 2. Physical exercise funded by the European Union’s Horizon 2020 research and 3. Nutrition innovation programme. The project aims to develop a personal 4. Medication Mobile application 5. Environment health system to help CHF patients manage their disease. CHF management patients have to take various medications, monitor their weight, 6. Calendars and plans exercise appropriately, watch what they eat and drink, and make other changes to their lifestyle. The HeartMan system will provide accurate advice on disease management adapted to each patient in Figure 1: Overall architecture of the HeartMan DSS. a friendly and supportive fashion. The DSS follows the best 67 The overall architecture of the HeartMan DSS is shown in Figure To date, all these modules have been designed in terms of their 1. The system will use wrist-band sensors to monitor patient’s functionality, requirements, input data, processing, outputs, and physical activity, heart rate and some other physiological signs. In distribution between the client (patient’s mobile app) and the addition, it will receive data from additional devices, such as server (the DSS in “the cloud”) [3]. scales, smartphone and from the patient via the user interface of the mobile application. This will allow the system to identify the 3. PHYSICAL EXERCISE MODULE patient's current physical and psychological characteristics. This The HeartMan DSS administers a comprehensive exercise data will be combined with patient’s health data to help them programme. At the beginning, the DSS collects medical decide on disease control measures. The advice will be tailored to information and assesses patient’s physical capacity in order to the patient's medical condition by adapting it to the patient's plan the difficulty level of the exercises. Then, the DSS provides a psychological profile (such as normal, poorly motivated, weekly set of endurance and resistance exercises, which increase depressed, and anxious) and current health state. The advice will in difficulty as the patient becomes fitter. The DSS also guides the be shown at the mobile app. A web-based interface will be patient during each exercise session: it checks whether the patient provided to the physician, too, who will be able to monitor the is ready to start, then provides instructions, and finally asks the patient’s health state and progress, and define or approve patient to evaluate the exercise. The exercise module follows the parameters that affect the advice given to the patient. guidelines provided in [5] with minor modifications to fit in a The core of the HeartMan DSS are modules that interpret mobile application. patient’s data and make recommendations. There are six main modules, which address the following aspects of health 3.1 Physical Capacity Assessment management: Prior to starting using the HeartMan DSS, the patients should perform a cardiopulmonary exercise (cycloergometry) test to 1. Physiological measurements: CHF patients should perform assess their physical capacity. Alternatively, when using the various physiological measurements on a regular basis, such system in a supervised, standardized setting, patients can perform as measuring their weight, blood pressure, heart rate, etc. This a 6-minute walking test. On this basis, the physical capacity of raises the patients’ awareness of their health, provides each patient is assessed as “low” (less than 1 W/kg measured by valuable information to their physicians, and provides inputs cycloergometry or less than 300 m walked in 6 minutes) or to the DSS. For this purpose, the DSS reminds the patient to “normal” (otherwise). regularly perform these measurements, provides the functionality to carry them out, and manage the collected data. 3.2 Weekly Exercise Planning The DSS system provides the patient with a combined endurance 2. Exercise: Physical conditioning by exercise training reduces and resistance exercise programme. Both types follow the same mortality and hospitalization, and improves exercise tolerance principle described with four parameters: frequency (times per and health-related quality of life. For this purpose, the DSS week), intensity, duration and type. These parameters are provides a comprehensive exercise programme, which is combined with the physical capacity to make a weekly exercise detailed later in section 3. plan for each patient. For instance, low-capacity patients start with 3. Nutrition: CHF patients should maintain their body weight very light 10-15-minute endurance exercises twice per week. and take care of their diet, for instance, not eating too much According to the patient’s progress, these parameters may change salt or drinking too much fluid. The DSS assesses the with time. In the HeartMan DSS, the progress is prescribed by patients’ nutrition behavior, educates them through a quiz and two models: provides advice towards a healthy diet.  EnduranceFrequency: a model for suggesting weekly 4. Medication: Good adherence to medication therapy decreases frequency of endurance exercises; mortality and morbidity, and improves well-being of CHF patients. For this purpose, the DSS reminds the patient to take  EnduranceTime: a model for suggesting weekly time medications and assesses the patient’s adherence to the boundaries of endurance exercises. medication scheme. For each medication, the patient may obtain an explanation why the adherence is important. Both models are formulated using a qualitative multi-criteria decision analysis method DEX [6]. Here, we illustrate the 5. Environment management: Environmental conditions, such as approach describing the EnduranceFrequency model, whose temperature and humidity, may affect the patient’s feeling of struc DEX tu i re is shown in Fig E u nd re 2 uranc . eFrequency.dxi 15.5.17 Page 1 health. Combining both, the patient’s and environmental Scales conditions, the DSS advises the patient how to change the Attribute Scale environment to improve their health feeling. EnduranceFrequency 2x; 3x; 4x; 5x Normative 2x; 3x; 4x; 5x Category low; normal 6. Calendars and plans: Given all the DSS aspects Week 1; 2; 3; 4; 5; 6; 7; 8; 9; 10; 11; 12; 13; 14; 15; 16; 17; 18; 19; 20; 21; 22; 23; 24; more Current 2x; 3x; 4x; 5x (measurement, exercise, nutrition, medication and Transition decrease; stay; increase; automatic MedicalAssessment decrease; stay; increase; automatic environment management) and many interactions between PatientsAssessment decrease; stay; increase; automatic them, it is important to sensibly arrange all the activities and Figure 2: Structure of the EnduranceFrequency model. notifications, for instance not sending nutrition advice during exercise or suggesting physical exercise after taking diuretics. The EnduranceFrequency model is aimed at suggesting the This DSS module thus coordinates the activities and arranges frequency of exercises for the next week, based on the patient’s all the plans into one single calendar. physical capacity, week in the programme, current frequency, and the possible physician’s and patient’s suggestions for the change. In other words, the model takes into account both the normative 68 (as proposed by a general programme) and actual (as practiced by exercise checking, (3) exercise monitoring, and (4) post-exercise the patient) frequency, leveraging the patient’s and physician’s assessment. opinion about the suggestion for the subsequent week. The overall recommendation, which is 2, 3, 4, or 5 times per 3.3.1 Reminding the patient week, is represented by the root attribute EnduranceFrequency Patients can choose the days when they want to exercise (e.g., (Figure 2). The recommendation depends on three sub-criteria: every Tuesday, Thursday and Sunday). On these particular days, the patients are at some predefined morning time reminded about 1. Normative: Frequency as suggested according to the default the daily exercise. Another reminder is issued if the exercise has programme. It depends on the patient’s physical capacity not been completed before the given afternoon time. (“low” or “normal” Category) and the current Week. The progression is defined by rules presented in Table 1. 3.3.2 Before the exercise 2. Current: The frequency of exercises currently carried out by Before the start of each exercise session, the HeartMan DSS the patient; it can run ahead or behind the Normative plan. In checks if all prior-exercise requirements are met, and advises the order to make only small and gradual changes to the patients about safety. Figure 3 shows the decision model. frequency, Current is compared to Normative and only a one- step change is suggested in each week. 3. Transition is an attribute that captures the patient’s wish and the physician’s opinion about changing the frequency. The possible values are “decrease”, “same”, “increase” or “automatic”; the latter is meant to suggest the frequency according to the normal plan, for instance, when neither the patient or physician have given any suggestion. The patient’s and physician’s suggestions are combined according to decision rules shown in Table 2. The first two rules say that whenever the patient or the physician suggest to decrease the frequency, it should indeed be decreased (the symbol ‘*’ represents any possible value). Rules 3 and 4 suggest to keep the current frequency whenever one of the participants suggests so, unless the other participant suggests “decrease” Rules 5 and 6 define a similar reasoning for “increase”. If both participants have no particular suggestions, the “automatic” transition according to the normal plan takes place. Figure 3: Pre-exercise assessment. Table 1: Decision table defining the Normative frequency. 1. Information requirements: The blood pressure should have been measured during the day. If not, the patients are Category Week Normative instructed to measure it. The pre-exercise heart rate is 1 low <=4 2x measured automatically by the wristband; the system makes 2 low 5–12 3x sure that it is actually worn. 3 normal <=6 3x 4 low 13–18 4x 2. Medication requirements: Patients are asked to fill a check-list 5 normal 7–12 4x of frequently seen side effects based on their medication 6 low >=19 5x schemes and symptoms (e.g., dizziness and chest pain). On 7 normal >=13 5x this basis the DSS checks for any possible restrictions due to medications or symptoms and suggests rescheduling the session if necessary. The physician or nurse are contacted if Table 2: Decision rules for Transition. severe side effects are present. In the case of dizziness or chest MedicalAssessment PatientsAssessment Transition pain, patients are instructed to rest until the symptoms are no 1 decrease * decrease longer present. 2 * decrease decrease 3. Physiological requirements: If all the requirements are met, 3 stay not decrease stay patients can start with the exercise, otherwise they are 4 not decrease stay stay instructed to repeat the measurements after five minutes of 5 increase increase or automatic increase rest. If after re-checking the measurements are still not within 6 increase or automatic increase increase safe limits, exercise is not allowed and patients are advised to 7 automatic automatic automatic contact their physician or heart failure nurse. Again, a DEX [3, 6] model is employed for assembling and 3.3 Daily Exercise Management checking the medical conditions, which include medical intake, comorbidities and current physical condition of the patient. The Once a weekly plan has been established, the HeartMan DSS structure and scales of the PreExerciseRequirements model are assists the patient in carrying out their daily exercises. This shown in Figure 4. All attributes are binary (“yes”/“no” or consists of four activities: (1) reminding the patient, (2) pre- “not_met”/“met”). The values of the input attributes are 69 determined from patient data whenever the pre-exercise 3.3.4 After the exercise requirements are checked (normally once per day before making After completing the exercise, the patients can rate their feeling of exercises). The subtrees of the model comprise four main groups intensity (very light, light, moderate, intense, very intense). Then of possible reasons against participating in the exercises: the system assesses the exercise based on measurements recorded during the exercise. It checks if the exercise was prematurely  Blood coagulation: Whenever the patient takes anticoagulants finished and if the intensity was on average in the prescribed and there are symptoms indicating a possible bleeding: rash, limits. The system takes into account this information when hemorrhages, or neurological symptoms. assessing the adherence to the exercise plan and the patient’s improvement. Independent of this, the exercise is shown as  Medication intake: Whenever one of the following completed and the weekly plan is updated. medications has been taken 2 hours or less before the exercise: beta blockers, ACE inhibitors, ARBs, diuretics, or loop diuretics. 4. CONCLUSION  This paper described the design of the HeartMan DSS that is Heart rate: Whenever the patient takes Digitalis and his/her concerned with “medical” interventions (i.e., interventions that try HR is less than 45 bpm. to improve the patients’ physical condition as opposed to  Blood pressure: Whenever there are risks of hypertension psychological). The DSS is based on clinical guidelines for the (taking ACE inhibitors or ARBs, and the patient has persistent self-management of CHF, additional medical literature and expert low blood pressure or persistent cough) or problems regarding knowledge from the project consortium. The DSS is designed in the systolic blood pressure (when the patient’s systolic blood terms of process models (the order of actions and questions) and pressure is less than 105 and he/she recently took loop decision models (how to make some complex decisions – diu DEXiretics). PreExerciseRequirements.dxi 16.5.17 branch P inag g e 1 in the process models) for five main topics of CHF Scales management: physiological measurements, exercise, nutrition, medication, and environment management. Attribute Scale PreExerciseRequirements not_met; met The DSS is currently being integrated in the overall HeartMan BloodCoagulationReasons yes; no TakesAnticoagulats yes; no platform. A comprehensive validation involving 120 CHF patients PossibleBleeding yes; no (of whom 40 are controls without the DSS) is planned for 2018. Rash yes; no Hemorrhages yes; no NeurologicalSymptoms yes; no MedicationIntakeReasons yes; no 5. ACKNOWLEDGMENTS Intake<2hours yes; no ExercisePreventionMedications yes; no The HeartMan project has received funding from the European TakesBetaBlockers yes; no Union’s Horizon 2020 research and innovation programme under TakesACEInhibitors yes; no TakesARBs yes; no grant agreement No 689660. Project partners are Jožef Stefan TakesDiuretics yes; no Institute, Sapienza University, Ghent University, National TakesLoopDiuretics yes; no HeartRateReasons yes; no Research Council, ATOS Spain SA, SenLab, KU Leuven, MEGA TakesDigitalis yes; no Electronics Ltd and European Heart Network. HR<45 yes; no BloodPressureReasons yes; no HypertensionReasons yes; no TakesACEInhibitors yes; no 6. REFERENCES TakesARBs yes; no PersistentLowBloodPressure yes; no [1] Heart failure. Health Information. Mayo Clinic. 23 PersistentCough yes; no December 2009. DS00061. http://www.mayoclinic.org/ SystolicPressureReasons yes; no TakesLoopDiuretics diseases-conditions/heart-failure/basics/definition/con- yes; no TookLoopDiuretics yes; no 20029801. SYS<105 yes; no [2] McMurray, J.J., Pfeffer. M.A. 2005. Heart failure. Lancet Figure 4: Structure of the PreExerciseRequirements model. 365 (9474): 1877–89. DOI= 3.3.3 During the exercise https://doi.org/10.1016%2FS0140-6736%2805%2966621-4. If the exercise is allowed, a list of exercises is shown to the [3] Bohanec, M., Vodopija, A., Dovgan, W., Maslov, P., patient, who can then select the preferred exercise. After selecting Luštrek, M., Puddu, P.E., Clays, E., Pardaens, S., Baert, A. the exercise, a detailed description (text or graphical) regarding (2017). Medical DSS for CHF. HeartMan Deliverable D4.1. the exercise is provided. [4] Turban, E., Sharda, R., Delen, D., King, D. (2010). Business Intelligence. 2nd Edition, Prentice Hall. During the exercise, the heart rate and systolic blood pressure are continuously measured by the wristband. The patients are advised [5] Piepoli, M.F., Conraads, V., Corrà, U., et al. (2011): to stop the exercise in case of symptoms or measurements lying Exercise training in heart failure: from theory to practice. A outside of prescribed safety margins. If the heart rate is within the consensus document of the Heart Failure Association and the safety limits, but too low or too high, the patent is advised to European Association for Cardiovascular Prevention and increase or decrease the intensity, respectively. The system also Rehabilitation. European Journal of Heart Failure 13, 347– advises the patients about the exercise duration and is capable of 357. recognizing a premature ending. [6] Bohanec, M., Rajkovič, V., Bratko, I., Zupan, B., Žnidaršič, M. (2013): DEX methodology: Three decades of qualitative multi-attribute modelling. Informatica 37, 49–5 70 Continuous Blood Pressure Estimation from PPG Signal Gašper Slapničar Matej Marinko dr. Mitja Luštrek Jožef Stefan Institute Faculty of Math. and Physics Jožef Stefan Institute Jamova cesta 39 Jadranska cesta 19 Jamova cesta 39 1000 Ljubljana 1000 Ljubljana 1000 Ljubljana gasper.slapnicar@ijs.si matejmarinko123@gmail.com mitja.lustrek@ijs.si ABSTRACT 2. RELATED WORK Given the importance of blood pressure (BP) as a direct Photoplethysmography is a relatively simple and affordable indicator of hypertension, regular monitoring is encouraged technique, which is becoming increasingly popular in weara- in people and mandatory for such patients. We propose bles for heart rate estimation. Exploring its applications, we an approach where photoplethysmogram (PPG) is recorded can see that it is also becoming more widely used in BP esti- using a wristband in a non-obtrusive way and subsequen- mation in one of two common approaches: 1.) BP estima- tly BP is estimated continuously, using regression methods tion from two sensors (PPG + Electrocardiogram (ECG)) based solely on PPG signal features. The approach is valida- or 2.) BP estimation using PPG only. ted using two distinct datasets, one from a hospital and the other collected during every-day activities. The best achie- PPG is based on illumination of the skin and measurement ved mean absolute errors (MAE) in a Leave-one-subject-out of changes in its light absorption. It requires a light source experiment with personalization are as low as 11.87 ± 12.31 (typically a light-emitting diode – LED light) to illuminate / 11.09 ± 9.99 for systolic BP and 5.64 ± 5.73 / 6.18 ± 4.85 the tissue (skin), and a photodetector (photodiode) to me- for diastolic BP. asure the amount of light either transmitted or reflected to the photodetector. Thus, PPG can be measured in either Keywords transmission or reflectance mode. With each cardiac cycle Photoplethysmography, blood pressure, regression analysis, the heart pumps blood towards the periphery of the body, m-health producing a periodic change in the amount of light that is absorbed or reflected from the skin, as the skin changes its tone based on the amount of blood in it [6]. 1. INTRODUCTION According to the World Health Organization (WHO), car- The first approach suggests the use of two sensors, typically diovascular diseases were the most common cause of death an ECG and a PPG sensor, in order to measure the time it in 2015, responsible for almost 15 million deaths combined takes for a single heart pulse to travel from the heart to a [1]. Hypertension is a common precursor of such diseases peripheral point in the body. This time is commonly known and can be easily detected with regular blood pressure (BP) as pulse transit time (PTT) or pulse arrival time (PAT) and measurements. its correlation with BP changes is well established [2]. Given the importance of BP, people should actively monitor The more recent approach is focused on PPG signal only, its changes. This is not trivial as the traditional ”gold stan- however the relationship between PPG and BP is only po- dard”BP measurement method involves an inflatable cuff, stulated and not well established, unlike the relationship which should be correctly placed directly above the main between PTT and BP. This approach is the least obtrusive artery in the upper arm area, at approximately heart height by far and PPG sensors have recently become very common [9]. These requirements impose relatively strict movement in most modern wristbands. restrictions on the patient and require substantial time com- mitment. Furthermore, when done by the patient himself, One of the earliest attempts at this approach was conducted the process can cause stress, which in turn influences the BP by Teng et. al. [3] in 2003. The relationship between arterial values, so it is most commonly done by medical personnel. blood pressure and certain features of the photoplethysmo- However, when BP is measured by medical personnel, this graphic (PPG) signals was analyzed. Data was obtained can again cause anxiety in the patient, commonly known as from 15 young healthy subjects in a highly controlled la- white coat syndrome. boratory environment, ensuring constant temperature, no movement and silence. The mean differences between the li- Our work focuses on analyzing the photoplethysmogram (PPG) near regression estimations and the measured BP were 0.21 and then developing a robust non-obtrusive method for con- mmHg for SBP and 0.02 mmHg for DBP. The correspon- tinuous BP estimation, which will be implemented and used ding standard deviations were 7.32 mmHg for SBP and 4.39 in an m-health system, based on a wristband with a PPG mmHg. sensor. A paper was published in 2013 in which authors used data 71 from Multiparameter Intelligent Monitoring in Intensive Care (MIMIC) waveform database [4] to extract 21 time domain features and use them as an input vector for artificial neural networks (ANNs). The results are not quite as good as the linear regression model described earlier, however the data is obtained from a higher number and variety of patients in a less controlled environment, but was still measured in a hospital setting and an undisclosed subsample of all ava- ilable data was taken. The results reached mean absolute difference between the estimation and the ground truth of less than 5 mmHg with standard deviation of less than 8 Figure 1: An example output of peak/cycle detec- mmHg [5]. tion algorithm on PPG signal. Black asterisks cor- respond to a detected peak while red circles corre- It is clear that the PPG only approach has potential, howe- spond to a detected cycle beginning. ver a robust method that works well on a general case is yet to be developed. 3.1.3 Cleaning based on ideal templates 3. METHODOLOGY In the second cleaning phase, a sliding window of 30 seconds The workflow consists of two main parts, namely the signal is taken and the mean of all cycles within this window is pre-processing and machine learning part. In signal pre- computed from the PPG signal. Presuming that the majo- processing, our PPG signal is cleaned of most noise and se- rity of cycles within a 30sec window are not morphologically gmented into cycles, where one cycle corresponds to a single altered, a good ”ideal cycle template” is created. Each indi- heart beat. Afterwards, features are extracted on per-cycle vidual cycle is then compared to this ideal template and its basis and fed into regression algorithms which build models quality is evaluated with three signal quality indices (SQIs). that are further evaluated. The most likely length of cycle L is always determined with autocorrelation analysis. The template is computed by al- 3.1 Signal pre-processing ways taking L samples of each cycle in the current window. When PPG is used in a wristband, the main problem comes from the contact between the sensor and the skin. During Signal quality indices are computed as follows. SQI1 : First everyday activity, the patient moves his arm a lot, which L samples of each cycle are taken, and each cycle is direc- in turn causes substantial movement artefacts in the signal. tly compared to the template using a correlation coefficient. This is partially alleviated by the usage of green light, which SQI2 : Each cycle is interpolated to length L and then the is less prone to artefacts, however pre-processing is still re- correlation coefficient is computed. SQI3 : The distance be- quired. tween template and cycle is computed using dynamic time warping (DTW). 3.1.1 Cleaning based on established medical crite- Finally thresholds for each SQI are determined and if more ria than half cycles in the given 30sec window are discarded, In first phase, both BP and PPG signal are roughly cleaned the whole window is considered too noisy and thus removed. based on established medical criteria [8]. During this phase, Example of this cleaning is shown in Figure 2. parts of signals with systolic BP (SBP) > 280mmHg or dia- stolic BP (DBP) < 20mmHg or the difference between SBP Once high quality signal is obtained, features can be extrac- and DBP < 20mmHg, are removed. This removes parts of ted from each cycle. signals for which the reference BP signal most likely con- tained an anomaly as such values indicate extreme medical 3.2 Machine learning condition and are not feasible in a common patient. In accordance with related work [5] several time domain fe- atures were computed and the set of features was further 3.1.2 Peak and cycle detection expanded with some from the frequency domain [8]. These In order to do further cleaning and feature extraction, PPG are shown in Figure 3. cycle detection is mandatory. This is again not trivial, as substantial noise in the PPG signal poses a significant pro- These features were extracted for each cycle and used in blem. machine learning to derive a regression model for BP esti- mation. A slope sum function, which enhances the abrupt upslopes of pulses in the PPG signal is first created. Afterwards, a time- 4. EXPERIMENTS AND RESULTS varying threshold for peak detection is applied [7]. After the peaks are detected, finding the cycle start-end indices is In an effort to make our method as general as possible, two rather simple as the valleys between peaks must be found. datasets were considered for our experiment and all the data An example of detected peaks and cycle locations is shown which had both PPG and BP signal were used. in Figure 1. 4.1 Data Once cycles are detected, they are used for further cleaning First is the publicly accessible MIMIC database set from and feature extraction. which all the patients having both PPG and arterial BP 72 Figure 2: An example of the cleaning algorithm in the 2nd phase. Comparing the top (uncleaned) and bottom (cleaned) signal, we see that the obvious artefact period is discarded. less accurate and extremely sensitive to body position. In the first phase of data collection, 8 healthy subjects were considered, 5 male and 3 female. Each wore the wristband for several hours during every-day activities and measured their BP every 30min or more often. Finally, only parts of signals 3 minutes before and after the BP measurement were taken into consideration. 4.2 Experimental setup Leave-one-subject-out experiment was conducted on each dataset, as it is the most suitable experiment to evaluate the generalization performance of the algorithms. Due to time and computational power restrictions, data was subsampled Figure 3: Time domain features which were used. by taking 500 uniformly selected cycles. Tc = cycle time, Ts = systolic rise time, Td = di- astolic fall time, AAC = area above the curve and During the initial attempt, a regression model was trained AUC = area under the curve for systolic and dia- in each iteration on all subjects, except the left out. This stolic part of a cycle. yielded poor results, hinting at the fact, that most patients are unique in some way. This was confirmed by doing a cycle morphology analysis during which it was established (ABP) signal were taken. This results in 50 anonymous that different subjects have different cycle shapes and that patients, each having on average several hours of both signals similar cycle shapes do not signify similar BP values. Thus, available. The data was collected in a hospital environment, personalization of the trained models was considered. using the hospital equipment. In the second attempt, the regression models were again Second is a dataset collected at Jozef Stefan Institute (JSI) trained using all subjects except the left out, however they using the Empatica E4 wristband for PPG and an Omron were further personalized using some data instances from cuff-based BP monitor for the ground truth BP, as is com- the left out subject. The instances of the left out subject mon in related work [3]. The collection procedure was con- were first grouped by BP values and these groups were then ducted in accordance with recommended clinical protocol sorted from lowest to highest BP. Afterwards, every n-th [9], ensuring correct placement of the cuff on upper arm group (n = 2,3,4,5,6 ) of instances was taken from the testing with the sensor above the main artery and its location at data and used in training in order to personalize the model approximately heart height. The subjects were in a sitting to the current patient. This ensures personalization with upright position during the measurements, thus following different BP values, as taking just a single group of instances the protocol as best as possible. Ideally, arterial BP would gives little information, since the BP will be constant within be measured in the artery as ground truth, however due to this group. Given the fact that MIMIC data consists of invasive nature of the procedure, this is not feasible in an roughly 5x the amount of patients compared to JSI collected everyday life situation, so an upper arm cuff-based digital data, the personalization data for it was multiplied 5 times, monitor was used as a good replacement. These devices are making it noticable within the large amount of training data superior to wrist cuff-based monitors, as wrist devices are from the remaining patients. 73 12.31 for SBP and 5.64 ± 5.73 for DBP. Due to high amount of movement artefacts in JSI collected data, a lot of data was removed by the cleaning algorithm, leaving a very low amount of usable data with very low va- riations in BP. This further enhanced the performance of dummy regressor, while leaving little information for other algorithms. Best achieved errors of 11.09 ± 9.99 for SBP and 6.18 ± 4.85 for DBP are only slightly surpassing the mean predictions at maximum personalization, as shown in Figure 5. 5. CONCLUSION We have developed a pipeline for BP estimation using PPG signal only and have evaluated its performance on two dis- Figure 4: MAE for SBP and DBP for MIMIC data- tinct datasets. set at different amounts of personalization. First part of the pipeline does signal pre-processing, remo- ving most movement artefacts and detecting PPG cycles. The second part computes features on per-cycle basis and feeds them in regression algorithms. These were evaluated on hospital collected MIMIC database data as well as field collected data at JSI using a wristband. Due to low variati- ons in subject’s BP and high variation in their PPG, there is limited information about the correlation between the two, however promising results were obtained with best achie- ved mean absolute errors (MAE) in a Leave-one-subject-out experiment with personalization as low as 11.87 ± 12.31 / 11.09 ± 9.99 for systolic BP and 5.64 ± 5.73 / 6.18 ± 4.85 for diastolic BP. Acknowledgement The HeartMan project has received funding from the Eu- Figure 5: MAE for SBP and DBP for JSI collected ropean Union’s Horizon 2020 research and innovation pro- dataset at different amounts of personalization. gramme under grant agreement No 689660. Project partners are Institut Jozef Stefan, Unversita Degli Studi Di Roma La Sapienza, Universiteit Gent, Consiglio Nazionale Delle Richerche, ATOS Spain SA, SenLab, KU Leuven, MEGA During both attempts, several regression algorithms were Elektronikka Oy and European Heart Network. considered, as given in Figures 4 and 5. Mean Absolute Er- ror (MAE) was used as a metric. All models were compared 6. REFERENCES with a dummy regressor, which always predicted the mean BP value of the same combination of general and persona- [1] The World Health Organization. “The top 10 causes of lization data as the other models used to train themselves. death”, 2015. Finally, the regressor with the lowest MAE was chosen. [2] Geddes et. al. “Pulse transit time as an indicator of arterial blood pressure”, 1981. [3] Teng et. al. “Continuous and noninvasive estimation of For successful personalization, the user should measure his arterial blood pressure using a photoplethysmographic PPG continuously and also make a few periodic measure- approach”, 2003. ments of his BP using a reliable commercial device. This [4] Goldberger et. al. “PhysioBank, PhysioToolkit, and allows the model to personalize to the user, learning from a PhysioNet: Components of a New Research Resource for small sample of his labeled data, thus improving its predic- Complex Physiologic Signals”, 2000. tive performance. [5] Lamonaca et. al. “Application of the Artificial Neural Network for blood pressure evaluation with smartphones”, 2013. 4.3 Results [6] Allen. “Photoplethysmography and its application in Due to low variations in BP, the dummy regressor often clinical physiological measurement”, 2007. performs relatively well, however for MIMIC data with more [7] Lázaro et. al. “Pulse Rate Variability Analysis for BP variation, some improvements have been made as shown Discrimination of Sleep-Apnea-Related Decreases in the in Figure 4. The JSI collected data has proven to be more Amplitude Fluctuations of Pulse Photoplethysmographic Signal in Children”, 2014. problematic, as there are only a low amount of different BP [8] Xing et. al. “Optical Blood Pressure Estimation with values in this phase of collection. Photoplethysmography and FFT-Based Neural Networks”, 2016. The lowest error using MIMIC data was achieved by using [9] Frese, E. M. and Fick, Ann and Sadowsky, H. S. “Blood the RandomForest regression algorithm, with the highest Pressure Measurement Guidelines for Physical amount of personalization. The achieved errors were 11.87± Therapists“, 2011 74 Recognizing Hand-Specific Activities with a Smartwatch Placed on Dominant or Non-dominant Wrist Božidara Cvetković, Vid Drobnič, Mitja Luštrek Jožef Stefan Institute Department of Intelligent Systems Jamova cesta 39 Ljubljana boza.cvetkovic@ijs.si ABSTRACT In this paper we analyze and evaluate a possibility to use In this paper we analyze the use of accelerometer-equipped accelerometer equipped smartwatch to recognize a large set smartwatch to recognize hand-specific activities. We start of hand-oriented activities. Since many activities have sim- with a large set of activities, and since many activities have ilar pattern we gradually group semantically similar activ- a similar acceleration pattern, we gradually group seman- ities into single activity group to find a tradeoff between tically similar activities to find a tradeoff between the ac- accuracy and semantically understandable activity groups. curacy on one hand, and semantically understandable and Additionally, we compare the activity recognition in terms useful activity groups on the other hand. Additionally, we of number of activities and accuracy when wearing a smart- compare the activity recognition in terms of the number of watch on dominant or non-dominant wrist. activities and accuracy when wearing a smartwatch on the dominant or non-dominant wrist. The preliminary results The paper is structured as follows. Section 2 presents the show that we can recognize up to seven groups of activities related work on activity recognition, Section 3 introduces with the dominant, and up to five activity groups with the the dataset and methods for preprocessing and training the non-dominant wrist. models. The evaluation results are present in Section 4 and Section 5 concludes the paper. Categories and Subject Descriptors D.3.3 [Human-centered computing]: Ubiquitous and mo- 2. RELATED WORK bile computing Pioneers in activity recognition research studied use of sin- gle or multiple accelerometers attached to different locations Keywords on the users body. Attal et al. [5] reviewed the research Activity recognition, wrist wearable, machine learning, ac- done until 2015 and proved that number of recognized ac- celerometers tivities increases with the number of sensors attached to the users body. Since using one or more dedicated accelerom- 1. INTRODUCTION eters was perceived as unpractical, the researchers started Activity recognition is an important module in person ori- using devices that most people already have or will have in ented intelligent systems, since most of the further reason- the future, such as smartphones and wristbands. ing or assistance to the user depends on the user’s current or past activity. This dependency is highly significant in Research on activity recognition with the smartphone mostly applications intended for the management of lifestyle and covers analysis of accelerometer signals without any knowl- sports activities [6], as well as chronic diseases such as dia- edge of its orientation, thus recognizing only small fraction betes or chronic heart failure (CHF). In diabetes, the user of activities (walking, running, rest, etc.) [11]. Martin et needs to monitor two particular activities, the eating (which al. [10] was first to take varying orientation and location into increases the blood glucose level) and exercise (which de- consideration. Their approach requires use of all available creases the blood glucose) [1] and in CHF it is important to smartphone sensors to estimate the location and normalize monitor the food intake (eating) as well as exercise in terms the orientation. In our recent research [8], we proposed a of its intensity and amount of rest [4]. real-time method that normalizes the orientation, detects the location and afterward uses a location specific machine- Due to importance of activity recognition and availability of learning model for activity recognition. accelerometer equipped wearables it is not surprising that the research area is very popular and partially also very ma- The research on activity recognition with wrist-worn de- ture. The maturity of the area is shown in the amount of vices has started with the accelerometer placed on a persons applications and wearable devices dedicated to activity mon- wrist [5]. Since this is the most comfortable placement of the itoring available on the market [2, 3]. However, these appli- sensor, the research became popular for recognizing sports cations and devices mostly recognize three activities (walk- activities [12] and common activities (sitting, standing, ly- ing, running and rest), which is insufficient for applications ing, walking, running) [7]. However, none of the research in which e.g., eating or any other hand-oriented activities focused on recognizing hand-specific activities (e.g., eating, are important. washing, hammering, etc.), which is the topic of this paper. 75 3. MATERIALS AND METHODS The raw acceleration and heart rate data is first segmented 3.1 Dataset into 2-second windows, each next overlapping by half of its Dataset contains data of 11 volunteers equipped with two size, from which we extract 90 acceleration features and 4 smartwatches with accelerometer and a heart rate sensor heart rate features. In brief, the raw acceleration data is (one on each wrist), performing a predefined scenario. Av- first filtered (low-pass and band-pass) to remove noise and erage accelerometer sampling rate was 48.2 Hz (±4.4) for gravity. The data is then used for calculation of physical the left hand and 51.3 Hz (±14.2) for the right hand. (e.g., velocity, kinetic energy, etc.), statistical features (e.g., the mean, variance, etc.) and features based on signal pro- The scenario contained 39 different activities, but not all cessing expert knowledge (e.g., number of peaks in a signal, were performed by each volunteer. Figure 1 presents the etc.). The reader is referred to [8] for more details about the distribution of data in terms of number of learning instances feature extraction. Once the features are extracted, they in the dataset and in terms of people performing the activity are used to form a feature vector to be used for machine- (see Section refpreprocess). We collected approximately two learning. hours of data per person. We can observe that some activi- ties were performed by one person only, which is insufficient 3.3 Method for training the models and evaluating them using leave-one- Activity recognition is set as a classification task, performed subject-out approach. Omitting these activities left us with in real-time. The feature vector formed during the feature 30 activities, due to errors in data collection we also had extraction (Section 3.2) is feed into a classification machine- to omit the mobile use, phone call, clapping, white board learning model trained to recognize the activities. and rolling dice. This left us with 25 activities for further analysis. The collected dataset contains data labeled with 25 activities for each wrist. To design an accurate classifier, we had to Number of Instances solve two challenges: (i) the difference in movement of domi- 0 500 1000 1500 2000 2500 3000 3500 1 Basketball nant and non-dominant hand during the same activity (e.g., 11 Brushing Teeth drinking, eating, writing, etc.), and (ii) similar hand move- 1 Cards ment when performing different activities. We decided to 10 Cleaning Surfaces 11 Computer Use develop two classification models, one for each wrist accord- 11 Cooking ing to dominance to solve the first challenge. For the second 1 Cycling 1 Digging challenge we analyzed the possibility to semantically group 8 Drinking the activities, thus achieve higher recognition accuracy but 9 Eating Cutlery still keep understandability of the recognized activity. 1 Eating Sandwich 3 Hammering 4 Origami To select the machine-learning classification algorithm to be 11 Reading 2 Rolling Dice used for training the models, we have first evaluated the 5 Rubik's Cube classification accuracy of five different machine-learning al- 0 Running gorithms as implemented in Weka suite [9] (J48, SVM, JRip, 2 Sawing 2 Tennis Random Fores and Na¨ıve Bayes) on the dataset with 25 ac- eople 2 Using Screwdriver tivities. All experiments are done with Leave-One-Subject- of P 10 Vacuuming 2 Volleyball Out approach (LOSO). As in our previous activity monitor- 11 Washing Face Number ing research, the Random Forest achieved the best results 11 Washing Hands and was chosen for all further experiments. 1 Weeding 11 Writing 2 Clapping Once the machine-learning algorithm was chosen we ana- 3 Drawing 6 Eating lyzed the possible grouping of the activities according to 3 Eating Apple dominance. We started with the dominant hand, the group- 2 LEGO ing of which is presented in Figure 2. We start by gradually 2 Mobile Use 2 Phone Call grouping the most similar activities together and evaluating 2 Chess the impact on accuracy. We first group the activities that 1 Stairs 1 Sweeping seem the most similar. All upper hand movements used 2 Walk in face hygiene are grouped together, next are the eating 2 White Board activities, sports and activities similar to writing. The fi- 2 Watering Flowers nal three groups are the activities where the person plays Right Left games or the hand gesture is of low intensity. We also tried to group home chores into low and high intensity, which Figure 1: Number of instances per activity (y axis) turned out less accurate then if grouping all home chores to- and number of people performing the activity (x gether. With this approach we divided all 25 activities into 7 axis) groups or classes to be recognized when smartwatch is worn on dominant hand. Results of each iteration is presented in 3.2 Preprocessing Section 4. The goal of the preprocessing procedure is to combine the accelerometer and heart rate data received from the smart- The same approach was used to group activities to be rec- watch into form suitable for further use with machine-learning ognized by non-dominant hand (Figure 3). We grouped the algorithms (feature vectors). sports activities, eating activities, all chores activities to- 76 Figure 2: Grouping of activities when smartwatch is Figure 3: Grouping of activities when smartwatch is worn on a dominant hand. worn on a non-dominant hand. gether. The activities that were left were very similar in terms of non-dominant hand movement. We tried to dis- tinguish between activities which are similar to writing and games activities, but this decreased the accuracy compared to grouping the two types of activities into single group (hand work). The last group of activities contains the wash- ing activities. With this approach we divided all 25 activities manner on dataset presented in Section 3.1. The results are into 5 groups or classes to be recognized when smartwatch presented in Table 1. is worn on non-dominant hand. The results of each iteration are presented in Section 4 First, we evaluated the use of acceleration and heart rate data retrieved from the smartwatch attached to the domi- Apart from evaluating the classification models for each wrist nant wrist. We used the default approach (D) introduced in on dedicated grouping of activities, we have also evaluated Section 3.3 to evaluate each grouping of the activities. The the use of non-dominant hand activities grouping for train- increase in accuracy while gradually decreasing the num- ing the dominant hand model and vice-versa (denoted as ber of recognized classes from 25 to seven is presented in Cross). Both experiments were preformed in two ways: Figure 4. As expected the accuracy increased with each subsequent grouping and we have finally settled for seven classes (Dominant wrist (D)). If we apply smoothing (Dom- • By using the machine-learning model trained for the inant wrist (S)) we gain 3 percentage points in accuracy. specific wrist directly, namely Default approach (D) Finally, we evaluated the recognition of seven classes with non-dominant wrist data, which returned poor accuracy of • By smoothing the results using the majority classifica- 58% when default (D) method was used and 63.7% when tion in the 10-class sliding window, namely smoothing smoothing was applied (Figure 4 Cross: Non-dominant wrist approach (S). The length of the window was selected (S)). arbitrarily. The same approach was used to define the classes to be recognized with non-dominant hand. We first used the de- The results are presented in Section 4. fault approach (D) on each grouping of the activities which resulted in five final classes. The process of grouping and 4. EVALUATION respective accuracy is presented in Figure 4 (Non-dominant The goal of the evaluation was to analyze and compare the wrist (D)). When smoothing is applied (Non-dominant wrist recognition of the activities according to the retrieved data (S)) we gain 4 percentage points in accuracy. Finally, we from the smartwatch worn on the dominant or non-dominant evaluated the recognition of five classes with dominant wrist wrist. Additionally, we wanted to evaluate and get an insight data, which as expected returned higher accuracy then with into type of activities that can be recognized in respect to seven classes (76% when default (D) method was used and the hand dominance. The evaluation was performed with 84% when smoothing was applied (Figure 4 Cross: Domi- Random Forest algorithm in leave-one-subject out (LOSO) nant wrist (S)). 77 Heart Network. We would also like to thank student Grega Table 1: Evaluation of activity recognition. The Mezič for help with the data collection. methods: D=default, S=smoothed. Wrist (method) Accuracy [%] # classes 7. REFERENCES Dominant (D) 71 7 [1] American Diabetes Association. Dominant (S) 79 7 http://www.diabetes.org/food-and-fitness/. Cross: Non-dominant (D) 58 7 [Online; accessed September-2017]. Cross: Non-dominant (S) 64 7 [2] FitBit. https://www.fitbit.com/eu/home. [Online; Non-Dominant (D) 70 5 accessed September-2017]. Non-Dominant (S) 74 5 [3] Runkeeper. https://runkeeper.com/. [Online; Cross: Dominant (D) 76 5 accessed September-2017]. Cross: Dominant (S) 84 5 [4] P. A. Ades, S. J. Keteyian, G. J. Balady, N. Houston-Miller, D. W. Kitzman, D. M. Mancini, and M. W. Rich. Cardiac Rehabilitation Exercise and Self-Care for Chronic Heart Failure, 2013. [5] F. Attal, S. Mohammed, M. Dedabrishvili, F. Chamroukhi, L. Oukhellou, and Y. Amirat. Physical Human Activity Recognition Using Wearable Sensors. Sensors (Basel, Switzerland), 15(12):31314–38, 2015. [6] S. Chatterjee and A. Price. Healthy Living with Persuasive Technologies: Framework, Issues, and Challenges. Journal of the American Medical Informatics Association, 16(2):171–178, 2009. [7] S. Chernbumroong and A. S. Atkins. Activity classification using a single wrist-worn accelerometer. 2011 5th International Conference on Software, Knowledge Information, Industrial Management and Applications (SKIMA) Proceedings, pages 1–6, 2011. [8] B. Cvetković, R. Szeklicki, V. Janko, P. Lutomski, and M. Luštrek. Real-time activity monitoring with a Figure 4: The grouping of activities and accuracy. wristband and a smartphone. Information Fusion, 2017. 5. CONCLUSION [9] M. Hall, E. Frank, G. Holmes, B. Pfahringer, We presented a feasibility study of recognizing hand-specific P. Reutemann, and I. H. Witten. The WEKA data activities using data retrieved from smartwatch on domi- mining software. SIGKDD Explorations Newsletter, nant or non-dominant wrist. We start with large set of 11(1):10, 2009. hand-specific activities and gradually decrease the number [10] H. Mart´ın, A. M. Bernardos, J. Iglesias, and J. R. of activities by semantically grouping them together. The Casar. Activity logging using lightweight classification preliminary results show that we can recognize larger set of techniques in mobile devices. Personal and Ubiquitous activity groups if we use data from the smartwatch worn on Computing, 17(4):675–695, 2013. dominant wrist ( 7 activity groups) then using data from the [11] M. Shoaib, S. Bosch, O. Incel, H. Scholten, and smartwatch worn on non-dominant wrist (5 activity groups). P. Havinga. A Survey of Online Activity Recognition Using Mobile Phones. Sensors, 15(1):2059–2085, 2015. Since these are only preliminary results, which gave us a fea- [12] P. Siirtola, P. Laurinen, E. Haapalainen, J. Röning, sibility insight, we will need to repeat the data collection pro- and H. Kinnunen. Clustering-based activity cedure to collect more samples of already recorded activities classification with a wrist-worn accelerometer using as well as record additional activities (e.g., sport-specific, basic features. In 2009 IEEE Symposium on home-chores specific, etc.). To achieve higher accuracy, we Computational Intelligence and Data Mining, CIDM will also need to perform feature selection procedure and 2009 - Proceedings, pages 95–100, 2009. analyze which features are relevant for the task. Finally, we will need to merge the dataset with other datasets that con- tain non-hand-specific activities and probably design more complex algorithm to achieve good results. 6. ACKNOWLEDGMENTS The HeartMan project has received funding from the Eu- ropean Union’s Horizon 2020 research and innovation pro- gramme under grant agreement No 689660. Project partners are Jožef Stefan Institute, Sapienza University, Ghent Uni- versity, National Research Council, ATOS Spain SA, Sen- Lab, KU Leuven, MEGA Electronics Ltd and European 78 R-R vs GSR – An inter-domain study for arousal recognition Martin Gjoreski Blagoj Mitrevski Mitja Luštrek, Matjaž Gams Department of Intelligent Systems, Faculty of Computer Science and Department of Intelligent Systems, Jožef Stefan Institute Engineering Jožef Stefan Institute Jožef Stefan International Skopje, R. Macedonia Jožef Stefan International Postgraduate School Postgraduate School Ljubljana, Slovenia Ljubljana, Slovenia martin.gjoreski@ijs.si ABSTRACT The affective states can be modeled using a discrete or a Arousal recognition is an important task in mobile health and continuous approach. In the discrete approach, the affect human-computer interaction (HCI). In mobile health, it can (emotions) is represented as discrete and distinct state, i.e., anger, contribute to timely detection and improved management of fear, sadness, happiness, boredom, disgust and neutral. In the mental health, e.g., depression and bipolar disorders, and in HCI continuous approach, the emotions are represented in 2D or 3D it can enhance user experience. However, which machine-learning space of activeness, valance and dominance [3]. Unlike the (ML) methods and which input is most suitable for arousal discrete approach, this model does not suffer from vague recognition, are challenging and open research questions, which definitions and fuzzy boundaries, and has been widely used in we analyze in this paper. affective studies [4] [5] [6]. The use of the same annotating model We present an inter-domain study for arousal recognition on six allows for an inter-study analysis. different datasets, recorded with twelve different hardware sensors In this study we examine arousal recognition from GSR and from which we analyze galvanic skin response (GSR) data from heart–related physiological data, captured via: chest-worn ECG GSR sensors and R-R data extracted from Electrocardiography and GSR sensors, finger-worn BVP sensor, and wrist-worn GSR (ECG) or blood volume pulse (BVP) sensors. The data belongs to sensor and pulse oximeter (PPG) sensor. The data belongs to six 191 different subjects and sums up to 150 hours of labelled data. publicly available datasets for affect recognition, in which there The six datasets are processed and translated into a common are 191 different subjects (70 females) and nearly 150 hours of spectro-temporal space, and features are extracted and fed into arousal-labelled data. ML algorithms to build models for arousal recognition. When one All of this introduces the problem of inter-domain learning, to model is built for each dataset, it turns out that whether the R-R, which ML techniques are sensitive. To overcome this problem, we GSR, or merged features yield the best results is domain (dataset) use preprocessing techniques to translate the datasets into a dependent. When all datasets are merged into one and used to common spectro-temporal space of R-R and GSR data. After the train and evaluate the models, the R-R models slightly preprocessing, R-R and GSR features are extracted and are fed outperformed the GSR models. into ML algorithms to build models for arousal recognition. Keywords Finally, the results between different experimental setups are compared, i.e., models that use only R-R features, models that Arousal recognition; GSR; R-R; machine learning; health. used only GSR features and models that use both R-R and GSR features. This comparison is performed in a dataset-specific setup 1. INTRODUCTION and merged setup where all datasets are merged in one. At the The field of affective computing [1] has been introduced almost end, the experimental results are discussed and the study is two decades ago and yet modeling affective states has remained a concluded with remarks for further work. challenging task. Its importance is mainly reflected in the domain of human-computer interaction (HCI) and mobile health. In the RELATED WORK HCI, it enables a more natural and emotionally intelligent Affect recognition is an established computer-science field, but interaction. In the mobile health, it contributes to the timely many remaining challenges. Many studies confirmed that affect detection and management of emotional and mental disorders recognition can be performed using speech analysis [21], video such as depression, bipolar disorders and posttraumatic stress analysis [8], or physiological sensors in combination with ML. disorder. For example, the cost of work-related depression in The majority of the methods that use physiological signals use Europe, was estimated to €617 billion annually in 2013. The total data from ECG, electroencephalogram (EEG), functional was made up of costs resulting from absenteeism and magnetic resonance imaging (fMRI), GSR, electrooculography presenteeism (€272 billion), loss of productivity (€242 billion), (EOG) and/or BVP sensors. health care costs of €63 billion and social welfare costs in the form of disability benefit payments (€39 billion) [2]. In general, the methods based on EEG data outperform the methods based on other data [4] [5], probably due to the fact the Affective states are complex states that results in psychological EEG provides a more direct channel to one’s mind. However, and physiological changes that influence behaving and thinking even though EEG achieves the best results, it is not applicable in [3]. These psycho-physiological changes can be captured by a normal everyday life. In contrast, affect recognition from R-R wearable device equipped with GSR, ECG or BVP sensor. For intervals or GSR data, is much more unobtrusive since this data example, the emotional state of fear usually initiates rapid can be extracted from ECG sensors, BVP sensors, PPG or GSR heartbeat, rapid breathing, sweating, and muscle tension, which sensors, most of which can be found in a wrist device (e.g., are physiological signs that can be captured using wearables. Empatica [9] and Microsoft Band [10]). Regarding the typical ML 79 approaches for affect recognition, Iacoviello et al. have combined 3. METHODS discrete wavelet transformation, principal component analysis and support vector machine (SVM) to build a hybrid classification 3.1 Pre-processing and feature extraction framework using EEG [11]. Khezri et al. used EEG combined with GSR to recognize six basic emotions via K-nearest neighbors 3.1.1 R-R data (KNN) classifiers [12]. Verma et al. [13] developed an ensemble The preprocessing is essential, since it allows merging of the six approach using EEG, electromyography (EMG), ECG, GSR, and different datasets. For the heart-related data, it translates the EOG. Mehmood and Lee used independent component analysis to physiological signals (ECG or BVP) to R-R intervals and extract emotional indicators from EEG, EMG, GSR, ECG, and performs temporal and spectral analysis. First, a peak detection (effective refractory period) ERP [14]. Mikuckas et al. [15] algorithm is applied to detect the R-R peaks. Next, temporal presented a HCI system for emotional state recognition that uses analysis, i.e., calculating the time distance between the detected spectro-temporal analysis only on R-R signals. More specifically, peaks, detects the R-R intervals. Once the R-R intervals are they focused on recognizing stressful states by means of the heart detected they can be analyzed as a time-series. First, each R-R rate variability (HRV) analysis. signal is filtered using median filter. After the median filter, However, a clear comparison between ML methods for affect person specific winsorization is performed with the threshold recognition from unobtrusively captured sensor data (e.g., R-R vs. parameter of 3 to remove outlier R-R intervals. From the filtered GSR data) has not been presented yet, since most of these studies R-R signals, periodogram is calculated using the Lomb-Scargle focused on only one dataset and a combination of the sensor data, algorithm [7]. The Lomb-Scargle algorithm is used for spectral aiming towards the highest performance and disregarding the analysis of unequally spaced data (as are the R-R intervals). obtrusiveness of the system. In this work, we analyze which ML Finally, the following HRV features were calculated from the time algorithms in combination with which data type (either R-R and spectral representation of the R-R signals: meanHR, meanRR, intervals or GSR) would yield best performance across six sdnn, sdsd, rmssd, pnn20, pnn50, sd1, sd2, sd1/sd2, lf, hf, lf/hf different datasets (domains) for arousal recognition. [29]. 2. DATA 3.1.2 GSR data To merge the GSR data, several problems were addressed. Each The data belongs to six publicly available datasets for affect dataset is recorded with different GSR hardware, thus the data can recognition: Ascertain, Deap, Driving workload dataset, Cognitive be presented in different units and different scales. To address this load dataset, Mahnob, and Amigos. Overall, nearly 150 hours of problem, each GSR signal was converted to µS (micro Siemens). arousal-labelled data that belong to 191 subjects. Table 1 presents Next, to address the inter-participant variability of the signal, the number of subjects per dataset, the mean age, number of trials person-specific min-max normalization was performed, i.e., each per subject, mean duration of each trial, duration of data per signal was scaled to [0, 1] using person specific winsorized subject - in seconds, and overall duration. minimum and maximum values. The winsorization parameter was Table 1. Experimental data summary. set to 3. Finally, the GSR signal was filtered using lowpass filter Duration per with a cut-off frequency of 1HZ. Dataset Subjects Females Mean age Trials trial [s] subject [min] dataset [h] The filtered GSR signal was used to calculate the following GSR Ascertain 58 21 31 36 80 48.0 46.4 features: mean, standard deviation, 1st and 3rd quartile (25th and DEAP 32 16 26.9 40 60 40.0 21.3 75th percentile), quartile deviation, derivative of the signal, sum of Driving 10 3 35.6 1 1800 30.0 5.0 the signal, number of responses in the signal, rate of responses in Cognitive 21 0 28 2 2400 80.0 28.0 the signal, sum of the responses, sum of positive derivative, Mahnob 30 17 26 40 80 53.3 26.7 Amigos 40 13 28 16 86 22.9 15.3 proportion of positive derivative, derivative of the tonic Overall 191 70 29.25 135 884.0 251.3 142.7 component of the signal, difference between the tonic component and the overall signal [21]. The four datasets, Ascertain, Deap, Mahnob and Amigos, were already labelled with the subjective arousal level. One difference between these datasets was the arousal scale used for annotating. 3.2 Machine learning For example, the Ascertain dataset used 7-point arousal scale, After the feature extraction, every data entry consists of 16 R-R whereas the Deap dataset used 9-point arousal scale (1 is very features and 14 GSR features, which can be input for typical ML low, and 9 is very high). From the both scales, we split the labels algorithms. Models were built using seven different ML in the middle, which is the same split used in the original studies. algorithms: Random Forest, Support Vector Machine, Gradient Similar step was performed for the Mahnob dataset. The two Boosting Classifier, and AdaBoost Classifier, KNN Classifier, datasets, Driving workload and Cognitive load, did not contain Gaussian Naive Bayes and Decision Tree Classifier. The labels for subjective arousal level. The Driving workload dataset algorithms were used as implemented in the Scikitlearn, the was labelled with subjective ratings for a workload during driving Python ML library [33]. For each algorithm, randomized search session. For this dataset, we presume that increased workload on hyper parameters was performed on the training data using 2- corresponds to increased arousal. Thus, we used the workload fold validation. ratings as an arousal ratings. The split for high arousal was put on 60%. Similarly, the cognitive load dataset was labelled for 4. EXPERIMENTAL RESULTS subjective stress level during stress inducing cognitive load tasks Two types of experiments were performed: dataset specific (mathematical equations). The subjective scale was from 0 to 4 experiments, and experiments with merged datasets. The (no stress, low, medium and high stress). We put the limit for high evaluation was performed using trial-specific 10-fold cross- arousal on 2 (medium stress). validation, i.e., the data segments that belong to one trial (e.g., 80 one affective stimuli), can either belong only to the training set or ML models. The results are presented in Figure 2. The results only to the test set, thus there was no overlapping between the show that the models that use the R-R intervals as input, training and test data. consistently outperform the models that use GSR features as input. 4.1 Dataset specific The results for the dataset specific experiments are presented in Table 2. The first column represents the ML algorithm, the second column represents the features used as input to the algorithm (R- R, GSR or Merged - M) and the rest of the columns represent the dataset which is used for training and evaluation using the trial- dependent 10-fold cross-validation. We report the mean accuracy ± the standard evaluation for the 10 folds. For each dataset, the best performing model(s) is(are) marked with green. For example, on the Ascertain and the Driving workload dataset, the best performing algorithm is the SVM, on the Deap dataset the best performing algorithm is the RF, on the Cognitive Load and the Mahnob datasets the best performing is the NB, and on the Amigos dataset the best performing is the AdaBoost algorithm. When we compare which input (R-R features, GSR features or Merged-M) provide better accuracy, on two datasets (the Asceratin and the Driving workload) the results are the same, on 5. CONCLUSION AND DISCUSSION the Deap dataset, the R-R features provide better results, on the We presented an inter-domain study for arousal recognition on six Cognitive Load dataset the highest accuracy is achieved both for different datasets, recorded with twelve different hardware the GSR and the Merged features, on the Mahnob dataset the GSR sensors. We experimented with dataset specific models and features provide best accuracy and on the Amigos dataset the models build on the overall (merged) data. We compared the Merged features. results of seven different ML algorithms, using three different feature inputs (R-R, GSR or Merged – M features). 4.2 Merged datasets For these experiments, all datasets were merged into one, and the trial-dependent 10-fold cross-validation was used to evaluate the Table 2. Dataset specific experimental results. Mean accuracy ± stdDev for trial-specific 10-fold cross validation. The best performing models per dataset are marked with green. Dataset Algorithm Features Ascertain Deap D. Workload Cog. Load Mahnob Amigos R-R 0.655 ± 0.07 0.556 ± 0.03 0.785 ± 0.24 0.739 ± 0.13 0.580 ± 0.11 0.536 ± 0.06 RF GSR 0.638 ± 0.06 0.503 ± 0.04 0.780 ± 0.24 0.763 ± 0.12 0.611 ± 0.07 0.473 ± 0.11 M 0.653 ± 0.05 0.540 ± 0.04 0.785 ± 0.25 0.755 ± 0.13 0.611 ± 0.10 0.559 ± 0.10 R-R 0.664 ± 0.07 0.536 ± 0.05 0.795 ± 0.26 0.717 ± 0.21 0.623 ± 0.15 0.521 ± 0.24 SVM GSR 0.664 ± 0.07 0.525 ± 0.05 0.795 ± 0.26 0.712 ± 0.20 0.588 ± 0.10 0.470 ± 0.12 M 0.664 ± 0.07 0.513 ± 0.03 0.795 ± 0.26 0.691 ± 0.18 0.623 ± 0.15 0.506 ± 0.13 R-R 0.649 ± 0.07 0.554 ± 0.03 0.785 ± 0.20 0.736 ± 0.15 0.578 ± 0.11 0.543 ± 0.06 GB GSR 0.642 ± 0.05 0.500 ± 0.04 0.800 ± 0.21 0.743 ± 0.12 0.609 ± 0.08 0.527 ± 0.09 M 0.644 ± 0.05 0.533 ± 0.03 0.755 ± 0.23 0.761 ± 0.15 0.609 ± 0.11 0.542 ± 0.09 R-R 0.658 ± 0.06 0.532 ± 0.02 0.750 ± 0.23 0.718 ± 0.13 0.580 ± 0.09 0.531 ± 0.07 AdaB GSR 0.633 ± 0.05 0.485 ± 0.03 0.750 ± 0.22 0.740 ± 0.13 0.589 ± 0.08 0.514 ± 0.09 M 0.623 ± 0.05 0.526 ± 0.03 0.755 ± 0.22 0.766 ± 0.16 0.610 ± 0.08 0.560 ± 0.08 R-R 0.625 ± 0.05 0.509 ± 0.02 0.710 ± 0.19 0.715 ± 0.13 0.582 ± 0.07 0.509 ± 0.05 KNN GSR 0.590 ± 0.06 0.496 ± 0.04 0.795 ± 0.26 0.772 ± 0.09 0.605 ± 0.06 0.533 ± 0.08 M 0.600 ± 0.05 0.490 ± 0.02 0.750 ± 0.23 0.770 ± 0.13 0.601 ± 0.09 0.533 ± 0.06 R-R 0.654 ± 0.07 0.537 ± 0.04 0.735 ± 0.15 0.748 ± 0.15 0.574 ± 0.06 0.485 ± 0.09 NB GSR 0.602 ± 0.04 0.537 ± 0.05 0.540 ± 0.22 0.803 ± 0.09 0.624 ± 0.07 0.454 ± 0.10 M 0.591 ± 0.04 0.535 ± 0.06 0.665 ± 0.17 0.804 ± 0.12 0.592 ± 0.06 0.486 ± 0.09 R-R 0.664 ± 0.07 0.519 ± 0.05 0.685 ± 0.17 0.736 ± 0.15 0.597 ± 0.09 0.505 ± 0.06 DT GSR 0.640 ± 0.05 0.542 ± 0.05 0.765 ± 0.22 0.734 ± 0.08 0.583 ± 0.09 0.483 ± 0.11 M 0.650 ± 0.05 0.524 ± 0.04 0.615 ± 0.22 0.704 ± 0.09 0.581 ± 0.13 0.551 ± 0.09 81 The results on the dataset specific setup showed that, out of the 12. M. Khezria, M.Firoozabadib, A. R. Sharafata. Reliable emotion ML algorithms tested, none yields the best performance on all recognition system based on dynamic adaptive fusion of forehead datasets. In addition to that, a clear conclusion cannot be made biopotentials and physiological signals. whether the R-R, GSR or the Merged features yield the best 13. G. K. Verma, U. S. Tiwary. Multimodal fusion framework: A results – this is domain (dataset) dependent. multiresolution approach for emotion classification and recognition from physiological signals. NeuroImage, 2014. On the merged dataset experiments, the R-R models slightly 14. R. M. Mehmooda, H. J. Leea. A novel feature extraction method outperformed the GSR models. This might be due to: (i) having based on late positive potential for emotion recognition in human more R-R features that GSR; (ii) having R-R features in frequency brain signal patterns. Computers & Electrical Engineering, 2016. domain but no GSR features in frequency domain; (iii) the 15. A. Mikuckas, I. Mikuckiene, A. Venckauskas, E. Kazanavicius2, R. method for merging the data from the heart-related sensors Lukas2, I. Plauska. Emotion Recognition in Human Computer providing more consistent features across datasets due to less Interaction Systems. Elektronika Ir Elektrotechnika, Reserarch noise in the ECG, BVP data. Journal, Kaunas University of Technology, 2014. In future, we plan to investigate intelligent combinations of ML 16. Wei Liu, Wei-Long Zheng, Bao-Liang Lu. Multimodal Emotion Recognition Using Multimodal Deep Learning. Online. Available at: models in order to gain accuracy. In addition to that, we plan to https://arxiv.org/abs/1602.08225, 2016. investigate more advanced techniques such as deep neural 17. W-L. Zheng, B-L Lu. A multimodal approach to estimating networks and transfer learning, which might be able to learn vigilance using EEG and forehead EOG. Journal of Neural general models that will be able to generalize across different Engineering, 2017. domains. Finally, once we find the best performing scenario, we 18. Z. Yin, M. Zhao, Y. Wang, J. Yang, J. Zhang. Recognition of will generalize the method for arousal recognition to method for emotions using multimodal physiological signals and an ensemble valence recognition and method for discrete emotion recognition. deep learning model. Comput Methods Programs Biomed. 2017. 19. K.Weiss, T. M. Khoshgoftaar, D. Wang. A survey of transfer 6. REFERENCES learning. Journal of Big Data, 2016. 1. R. Picard. Affective Computing. Cambridge, MA: MIT Press, 1997. 20. S. Schneegass, B. Pfleging, N. Broy, A. Schmidt, Frederik Heinrich. 2. Depression cost: A Data Set of Real World Driving to Assess Driver Workload. 5th http://ec.europa.eu/health//sites/health/files/mental_health/docs/matr International Conference on Automotive User Interfaces and ix_economic_analysis_mh_promotion_en.pdf, [Accessed Interactive Vehicular Applications, 2013. 27.03.2017]. 21. M. Gjoreski, M. Luštrek, M. Gams, H. Gjoreski. Monitoring stress 3. J. A. Russell. A circumplex model of affect. Journal of Personality with a wrist device using context. Journal of Biomedical and Social Psychology, 1980. Informatics, 2017, in press. 4. R. Subramanian, J. Wache, M. Abadi, R. Vieriu, S. Winkler, N 22. M. Gjoreski, H. Gjoreski, M. Luštrek, M. Gams. Continuous stress Sebe. ASCERTAIN: Emotion and Personality Recognition using detection using a wrist device: in laboratory and real life. ACM Commercial Sensors. IEEE Transactions on Affective Computing. Conf. on Ubiquitous Computing, Workshop on mentalhealth, pp. 2016. 1185-1193, 2016. 5. S. Koelstra, C. Muehl, M. Soleymani, J.-S. Lee, A. Yazdani, T. 23. M. Soleymani, T.Pun. A Multimodal Database for Affect Ebrahimi, T. Pun, A. Nijholt, I. Patras. DEAP: A Database for Recognition and Implicit Tagging, IEEE Transactions On Affective Emotion Analysis using Physiological Signals (PDF). IEEE Computing, 2012. Transaction on Affective Computing, 2012. 24. L. H. Negri. Peak detection algorithm. Python Implementation. 6. M.K. Abadi, R. Subramanian, S. M. Kia, P. Avesani, I. Patras. Online. Available at: http://pythonhosted.org/PeakUtils/. Nicu Sebe. DECAF: MEG-Based Multimodal Database for 25. M. Wu, PhD thesis. Michigan State University; 2006. Trimmed and Decoding Affective Physiological Responses. IEEE Transactions on Winsorized Eestimators. Affective Computing, 2015. 26. J.D. Scargle. Studies in astronomical time series analysis. II - 7. N.R. Lomb. Least-squares frequency analysis of unequally spaced Statistical aspects of spectral analysis of unevenly spaced data. The data. Astrophysics and Space Science, vol 39, pp. 447-462, 1976 Astrophysical Journal, vol 263, pp. 835-853, 1982. 8. I. Abdic, L. Fridman, D. McDuff, E. Marchi, B. Reimer, Schuller, 27. D. P. Kingma, J. Ba. Adam: A Method for Stochastic Optimization, B. Driver Frustration Detection From Audio and Video. Proceedings http://arxiv.org/abs/1412.6980, 2014. of the International Joint Conference on Artificial Intelligence 28. Tensorflow. Online. Available at: https://www.tensorflow.org/ (IJCAI'16), 2016. 29. R. Castaldoa, P. Melillob, U. Bracalec, M. Casertaa,c, M. Triassic, 9. M. Garbarino, M. Lai, D. Bender, R. W. Picard, S. Tognett. L. Pecchiaa. Acute mental stress assessment via short term HRV Empatica E3 - A wearable wireless multi-sensor device for real-time analysis in healthy adults: A systematic review with meta-analysis. computerized biofeedback and data acquisition. 4th International Biomedical Signal Processing and Control. 2015. Conference on Wireless Mobile Communication and Healthcare, pp. 3-6, 2014. 30. Scikit-learn, Python machine-learning library http://scikit- learn.org/dev/_downloads/scikit-learn-docs.pdf 10. Microsoft band. https://www.microsoft.com/microsoft-band/en-us 31. L.J.P, van der Maaten., G.E. Hinton. Visualizing High-Dimensional 11. D. Iacovielloa, A. Petraccab, M. Spezialettib, G. Placidib. A real- Data Using t-SNE. Journal of Machine Learning Research. 9: 2579– time classification algorithm for EEG-based BCI driven by self- 2605, 2008. induced emotions. Computer Methods and Programs in Biomedicine, 2015. 82 Predicting Office’s Ambient Parameters Vito Janko Andreas R. Stensbye Mitja Luštrek Jožef Stefan Institute Jožef Stefan Institute Jožef Stefan Institute Jožef Stefan International Jamova 39, 1000 Ljubljana Jožef Stefan International Postgraduate School andreas.stensbye@hotmail.com Postgraduate School Jamova 39, 1000 Ljubljana Jamova 39, 1000 Ljubljana vito.janko@ijs.si mitja.lustrek@ijs.si ABSTRACT possible actions that an user can take, and then recommend Bad environmental conditions in the office can negatively the one leading to the best working conditions. The system affect the workplace productivity. In the presented work we is meant to be used in offices without automatic ambient measure three ambient parameters - CO2, temperature and control and is a part of the larger ”Fit4Work” project [2] humidity - asses their quality and predict their likely future that is focused on helping to raise the well-being of office values. To do so, we first heuristically determine the state workers. It requires no prior knowledge or manual input of of the office (are the windows open, air conditioner active office properties, yet it is able to adapt to them over time. etc.) and then try to mathematically model the parameter’s future behavior. Based on the current and predicted state The ambient parameters were measured using the Netatmo of ambient parameters, we can send a recommendation on commercial device [1]. The same device is expected to be how to best improve them. Experimental evaluation shows used by the end users of the system, although it can be re- that our models outperform the related work in terms of placed with any similar device with the same functionality. prediction accuracy. This device has an indoor and an outdoor unit, both capa- ble of measuring the CO2, temperature and humidity, and Categories and Subject Descriptors sending the data to a web server. For easier testing and val- H.4 [Information Systems Applications]: Miscellaneous; idation of our method we also had sensors that monitor if D.2.8 [Software Engineering]: Metrics—complexity mea- the windows are opened and closed and an application where sures, performance measures we manually labeled the number of people in the office, the air conditioner state, heater state and humidifier state. As Keywords for the time of writing this paper we collected roughly two years of data for three different offices in our department. CO2, temperature, humidity, modeling, recommendations Data is continuously sent to a web server, where it is an- alyzed as described in Section 2. If a recommendation is 1. INTRODUCTION deemed necessary, it is sent to a mobile device via a push Good work environment is essential for keeping work pro- notification. ductivity. In this paper, we are focusing on three office’s ambient parameters: CO2, temperature and humidity. The The paper was inspired by another work [3], and proposes quality of these parameters is often hard for humans to ob- a different solution to the same problem. The proposed jectively detect, especially if they are changing slowly. How- solution makes heavier use of mathematical modeling and ever, it has been shown [4, 5] that when their quality drops produces more accurate predictions about the ambient pa- below certain thresholds, the work productivity in the office rameter’s future values. is negatively affected. 2. METHODOLOGY In this paper we present an intelligent system that is able to The goals of this paper are three-fold. First, to predict the measure these parameters and estimate their future values. state of the office: are the windows open, is air conditioner In the case of CO2 and temperature, a simple mathematical turned on, etc. This not only allows us to predict the prob- model is used for prediction, in the case of humidity a ma- able changes in the ambient parameters, but also to make chine learning model is used instead. Furthermore, it is able sensible recommendations: no need to recommend opening to asses the quality of these parameters, simulate several of the windows if they are already open. Second, to predict the future behavior of the following three parameters - CO2, temperature and humidity. We are interested in predicting values up to 30 minutes in advance. Our data was measured every 10 minutes, so this corresponds to 3 data points. Be- havior should be predicted for the current state and also for the cases where some office parameters changes. Finally, we use a combination of the previous two points to form rec- ommendations to the user on actions that improve the work environment. 83 While the physical phenomena of temperature, humidity to equalize itself with its surroundings and since all other and CO2 was already heavily studied in the past, the chal- factors (people, computers, etc..) only serve to warm the lenge we face here is in not knowing any attributes of the office, it is reasonable to conclude that such a temperature office where this system would be used: how big the of- drop was caused by the air conditioner. After a while of fice, how good the thermal insulation, the surface of the the air conditioner working, the temperature will converge windows, etc. Using standard formulas for predicting the to value that can be stored for later predictions. If the tem- ambient parameters can therefore be infeasible, given how perature starts rising again, the air conditioner is assumed many unknowns they contain. In our approach we tried to to be turned off. simplify the models to simple versions with only a few un- knowns. We use data recorded in the target office in the last 2.1.3 Heater two weeks (exact number of days may vary based on office The same assumptions and methods are used here as with usage) to estimate these unknowns, and then use them for the air conditioner, except in reverse: the heater is on if real-time predictions in the following day. the inside temperature rises significantly more then expected from the outside temperature, etc. 2.1 Virtual sensors Virtual sensors refer to values that are not directly mea- 2.2 CO2 predictions sured. Instead, their value is derived from the measured We start by modeling CO data and then later used to help derive some other value. In 2, as it the most ”well-behaved” of the three ambient parameters, and we describe the process in our setting, there are five virtual sensors that affect the am- depth. We later use a similar methodology for temperature bient parameters: the windows state, air conditioner state, modeling. Intuitively, CO heater state, humidifier state and number of people in the 2 level inside the office is increas- ing linearly with respect to the number of people present, office. but at the same time it tries to equalize itself with the out- side CO ”The number of people in the office” is calculated from rais- 2 level. The bigger the difference between outside and inside, the faster it moves from one side to another. If ing CO2 levels, and the humidifier state is tied to humidity window is opened, the same happens, only to a significantly data, so those two will be explored in the corresponding Sec- larger degree. This can be encapsulated in the following tions 2.2 and 2.4. The remaining three can be determined equation. with simple heuristics as described below. Cn+1 = Cn + α(Cout − Cn) + βp (1) 2.1.1 Windows Cn = CO2 inside at timestep n Windows were modeled in a binary fashion: they are ei- Cout = CO2 outside ther open or they are closed. In a real office there might p = the number of people in the room be many windows, some of them open, some closed, some α = the coefficient of diffusion speed (between 0 and 1) - perhaps half-open at any time; but lacking any knowledge small for closed windows, big for open ones about the window quantity or size, predicting their state β = how much a single person raises CO2 in a given time unit more accurately is almost impossible. An effect of opening the window is reflected on all three am- Using all the labeled data, the α and β are mostly trivial to bient parameters, but only in the case of CO compute using linear regression. Using them results in an 2 is the effect consistent. Whenever a window is opened, CO almost perfect match between the predicted and real values. 2 falls drasti- cally, whenever it is closed it starts to rise again. This allows In Figure 1 we plot a scenario where we know the initial CO2 us to make a simple heuristic: a.) if the CO level and all future windows states and all future numbers 2 is falling faster then some threshold, window was opened; b.) if CO of people, and we are able to predict CO2 level two days 2 keeps increasing, the window is closed; c.) if neither of those is in advance. This strongly signifies that the model captures happening, assume the last known state. Thresholds can the real-life behavior of CO2, and it is only a matter of be determined by looking at the data history and find such determining the correct coefficients. values that would generate predictions, where windows is opened/closed few times a day, as would realistically be the Calculating the coefficients for a given office without the la- case. beled data, however, is a challenging task as the above for- mula has 5 unknowns - α when windows are closed, α when This approach could be improved by correlating changes windows are opened, window state, β and number of people in CO p. Furthermore these coefficients can behave very similarly: 2 to those in temperature and humidity, but the de- scribed simple heuristic appeared to work well in practice. CO2 level in a room with many people and open window can be close to CO2 level in a room with closed windows and few people. The first improvement is to combine the 2.1.2 Air conditioner two variables β and p into one - γ, as we never need those Again we assume binary outcome - the air conditioner is ei- two individually and are only interested in their product. ther on or off - additionally we assume that the temperature This shortens the formula to: set on it is constant, or at least is changing infrequently. The distinguishing pattern of air conditioning is one of tem- perature inside decreasing while the temperature outside is higher then inside. Since the temperature naturally tries Cn+1 = Cn + α(Cout − Cn) + γ (2) 84 This formula can be rewritten in an analytical way (Equa- 2.4 Humidity predictions tion 3) so it can predict an arbitrary time step instead of Humidity was not changing much in our data, and when it only steps of integer size (10 minutes). A simple explana- did, there was no obvious pattern. So instead of plugging tion of this formula goes as follows - CO2 always converges the data into the same equation, we used a classical ma- to a value L. The number of people in the office dictates chine learning approach. The last few humidity and temper- this limit, while the value α dictates how fast we approach ature measurements, together with the window state and es- this limit. The inverse of this formula will also be useful and timated number of people (computed from γ in CO2 model) can be trivially computed using some basic algebra. are fed into a machine learning model, and a prediction for future humidity is given. Again the training of the model is made on the previous two weeks. If it turns out that ( the prediction underestimated the humidity in the office, γn, if α = 0 C the humidifier is determined to be active. If the classifier n = (3) L + (C0 − L)(1 − α)n, otherwise overestimates the humidity and humidifier was considered active, it is considered inactive from then on. γ L = Cout + α 2.5 Recommendation system Each ambient parameter has predefined quality ranges - Determining the window state is described in Section 2.1.1. good, medium and bad. For example: ”good” CO2 is under If we know the α value for the current window state, the γ 500 ppm, ”bad” over 800 ppm and ”medium” in between. value becomes the only unknown in the formula and can be The ideal case is to have all three parameters in the ”good” determined with a simple linear regression, using last three quality range. This, however, is not always possible as im- data points. Since γ correlates with the number of people in proving one parameter may damage another - opening the the office, it must be recalculated for every prediction. The window may improve the CO2, but it may reduce the tem- α value on the other hand is dependent on the office heat perature quality. The priority of the system is to have the insulation level, office size and windows size, and is therefore minimum number of ”bad” parameters. If all the parame- a constant. We can therefore estimate the α value by trying ters are ”medium” or above, the maximum number of ”good” different values on the past two weeks of data and then select parameters is prioritized. the one that has the lowest error rate when predicting - this is possible since when predicting on the past data, we already A possible action is a change in one of the devices/windows know what CO2 value will be reached. that exist in office. In the current version, all the devices are assumed to be binary (air conditioner is either on or off, windows are opened or closed, etc.). The list of all possi- ble actions is generated based on the current assumed state of the office. If the windows are assumed opened, ”open the 2.3 Temperature predictions window” action will be omitted. Some hand-selected actions We used the same base formula - Equation 3 - for the in- may appear in pairs, as they are commonly done simultane- side temperature prediction. This model however, has to be ously: turn on the air conditioner and close the windows for made more complex because of two factors. example. A default action ”do nothing” is also included on the list. First, the temperature does not converge towards the outside one, but goes towards some function of the outside temper- Each action effect is simulated over the period of 30 minutes. ature instead. For example, even if the outside temperature The action that results in the best state after that time is below zero, the temperature in the office never went below interval is selected. If the action has a higher score than the 10 degrees, even without heating. There are several reasons default action of doing nothing, it is recommended to the for this behavior, including the heat of the building itself, user. and the fact that building is warming and cooling at differ- ent rates than the exterior when the external temperature While not fully implemented yet, there are two areas with changes. This is dealt by calculating a function from last possible improvements that are currently worked on. One two weeks of data that models the expected inside temper- is to try to make the recommendations more time-specific. ature as a linear function of the outside temperature. The Instead of ”open the window”, we could recommend ”open calculation is made during rest days, when no one is in the the window for 7 minutes, then close again”. This can be office, reducing the noise in the data. This calculated value done by first determining all the relevant time frames - times then replaces the value Cout in the Equation 3. where a parameter shifts from one quality range to another. All the possible actions can then be tested against every rel- Second, we have to account for both air conditioning and evant time frame. Second is to predetermine which actions heating. The detection of their state is described in Sections are even sensible, given the context. If the only problem is 2.1.2 and 2.1.3. In the same section it is also described the temperature inside being too cold and it is also cold out- how to collect the limiting temperature value these devices side, then the sensible options are only to close the window generate. If either device is on, the corresponding limiting or to turn on the heater. This is being implemented by an value replaces L in Equation 3. Improvement of this rather ontology that contains facts about some ambient parame- simplistic modeling of the devices is subject to future work. ters, configured in a way that is able to search for relevant A prediction example is plotted in Figure 2. actions given current state. 85 Figure 1: CO2 prediction and ground truth, predicting values for the next two days, supposing that we have perfect information about the current and future office state. 4. CONCLUSIONS In this work we model three ambient parameters in the office. For two of them, we show a simple mathematical model that predicts their future behavior. For those two we get more accurate predictions than those in the related work. This is probably a consequence of using a physically-inspired for- mula. For humidity we use a machine learning model that while showing promising results, still has room for improve- ment. We also predict the state of devices and windows in the office, although the accuracy of this prediction has not yet been directly tested. Furthermore we presented a rec- ommendation system that we plan to test with multiple real offices in the future. 5. REFERENCES [1] ”netatmo. 2016. netatmo. (2016)”. ”https://www.netatmo.com”. ”Accessed: 2017-09-03”. Figure 2: Temperature prediction and ground truth. [2] B. Cvetković, M. Gjoreski, M. Frešer, M. Kosiedowski, We predict what happens if no action is done, and M. Luštrek. Monitoring and management of against what happens if window is opened. The pre- physical, mental and environmental stress at work. diction starts in the past so we can compare it to the [3] M. Frešer, A. Gradišek, B. Cvetković, and M. Luštrek. actual measurements. An intelligent system to improve thc parameters at the workplace. In Proceedings of the 2016 ACM 3. RESULTS International Joint Conference on Pervasive and As results we list (Table 1) the mean absolute error when Ubiquitous Computing: Adjunct, pages 61–64. ACM, predicting a parameter 30 minutes in advance, during a three 2016. month period. The test are made to be comparable with [4] L. Lan, P. Wargocki, and Z. Lian. Optimal thermal those in paper by Frešer et al. [3]. We show that our pre- environment improves performance of office work. dictions for CO REHVA European HVAC Journal, 2:12–17, 2012. 2 and temperature display lower error then the before-mentioned work. Their humidity measurements [5] D. P. Wyon. The effects of indoor air quality on were better, probably because of better selection of features performance and productivity. Indoor air, in their model. 14(s7):92–101, 2004. Table 1: Mean absolute error Parameter Our error Error reported by Frešer [3] CO2 [ppm] 43 79 Temperature [°C] 0.36 0.50 Humidity [%] 1.2 0.74 86 Real-time Content Optimization in Digital Advertisement Tom Vodopivec Davor Sluga Nejc Ilc University of Ljubljana University of Ljubljana University of Ljubljana Faculty of Computer and Faculty of Computer and Faculty of Computer and Information Science Information Science Information Science tom.vodopivec@fri.uni- davor.sluga@fri.uni-lj.si nejc.ilc@fri.uni-lj.si lj.si Gregor Sušelj Rok Piltaver Domen Košir Celtra Celtra Celtra Data Insights Team Engineering Analytics Team Data Insights Team gregor.suselj@celtra.com rok.piltaver@celtra.com domen.kosir@celtra.com ABSTRACT the target audience. Some research regarding optimization A key goal of advertising industry is to present the target of digital advertisements has been published [5, 7], however audience with advertisements that induce most engagement. a lot remains hidden from the scientific literature due to In digital advertising we are able to collect huge amounts of confidentiality agreements in the advertising industry. data on how different advertisements perform. This data can be used to optimize the content of such advertisements Advertisement content optimization can be done for long in real-time. The idea behind the optimization is essentially time spans – the decision on the content used in the ad- the same as in the multi-armed bandit problem. There are vertisements for the next campaign is made based on the many optimization algorithms available for solving it, but engagement rates in the previous campaign. On the other they need to be modified for the specifics of digital advertis- hand, optimization can also be done in real-time – the deci- ing. In this study, we analyse real data from hundreds of ad- sion on the content is made when an advertisement is about vertising campaigns and present a methodology to asses the to be shown on a web page or in a mobile application. In this potential of advertisement content optimization. We com- pare the performance of different optimization algorithms and propose their modifications. We conclude that only a small part of the advertising campaigns can potentially ben- efit from content optimization. However, when there is room for improving the performance of an advertisement, the op- timization algorithms coupled with the proposed modifica- tions are able to exploit most of it. Keywords Digital advertising, optimization, multi-armed bandit, se- lection bias, exploration-exploitation trade-off, Thompson sampling, Upper confidence bounds algorithm 1. INTRODUCTION Figure 1: Which advertisement is more engaging? In the advertising industry, there is a persistent aspiration to create and present to the target audience the most relevant paper, we are focusing on the latter approach. Our primary advertisement, which would produce the highest number of goal is to estimate the potential of the real-time content op- engagements from the viewers. The most frequently measure timization, i.e. the differences in engagement rates among of engagement, specific to the digital advertising, is the click different variants of an advertisement. If there is some po- rate. It measures how many views of a digital advertisement tential present, then the goal is to find the best algorithms led to a user clicking on the advertisement and thus showing that optimize the content of advertisements in real-time. interest in the advertised product or service. 2. CHALLENGES IN ADVERTISEMENT Given the complexity of the industry and the target popula- tion, it is difficult to predict which advertisement or variant CONTENT OPTIMIZATION of an advertisement would be more engaging. Fig. 1 shows A digital advertising campaign usually consists of different an example of two variants of an advertisement. Digital ad- variants of an advertisement, and generally multiple pos- vertising enables collecting large amounts of data that can sible media (different web pages and mobile applications) be used to measure the performance of advertisements. Us- where an advertisement is displayed. The variants of an ing an appropriate comparison methodology one can then advertisement can differ in many ways (e.g. text, graphics, decide which advertisement variant should be presented to colour). As the campaign progresses, one can track the num- 87 140000 1.20% 120000 1.00% 100000 s 0.80% re 80000 uos ss rate 0.60% 60000 Variant 1 Variant 1 Exp 0.40% 40000 Variant 2 Succe Variant 2 20000 0.20% 0 0.00% 0 50 100 150 200 250 300 0 50 100 150 200 250 300 Hours since campaign start Hours since campaign start (a) (b) Figure 2: Sample time lapse of exposures (a) and success rates (b) for an advertisement campaign with two variants. ber of exposures (i.e., how many times it was displayed to by the optimization algorithm and Sr is the number of suc- a viewer) and the number of engagements or successes (i.e., cesses achieved by displaying the variants randomly accord- the number of clicks on the advertisement) of all variants on ing to the uniform distribution. We usually express lift in a time basis (we collected hourly data). Our hypothesis is percentages. A positive lift means that using the optimiza- that this data can be used to identify the better performing tion algorithm is better than choosing variants randomly, variants and adjust the amount of times they are displayed whereas a negative lift means that the optimization has a in their favour. In practice, this is not a trivial task. The negative effect on the success rate. We estimate the poten- number of successes is usually very low in comparison to the tial of a campaign by computing the lift for the oracle algo- number of exposures, yielding near-zero probabilities of suc- rithm, which is the optimal algorithm that always (in each cess. Additionally, the data is non-stationary as shown on hour) chooses the best performing variant. Of course, the an example campaign in Fig. 2: the volume of exposures and oracle algorithm can only be used on the past data (not in the probability of success is changing over time, because of real-time), when it already knows how variants performed. hourly fluctuations in advertisement traffic, audience vari- ability, and seasonal trends. Therefore, large amounts of To cope with the large variance of the success rates, we use data are needed to obtain reliable estimates of each vari- one real campaign to generate several artificial campaigns by ants’ performance. Obtaining additional data is usually not applying different amounts of smoothing to compute hourly feasible, since exposures cost money and there are limited success rates for each variant. Only the uniformly selected funds available for a campaign. Another problem is that variants are used in creating the artificial campaigns to avoid any optimization done during the campaign introduces a se- selection bias affecting the success rates. Smoothing ranges lection bias [12] into the data – the variants chosen by the from producing completely stationary data on one hand to optimization algorithm have positively-skewed success rates. highly volatile trends (no smoothing) on the other hand. As a consequence, performance of certain variants can not From these campaigns we choose the best-case and the worst- be reliably estimated due to low number of exposures. case campaign according to its potential. This gives us the upper and lower limit for the true potential for different sce- When optimizing the advertisement content, one has to care- narios, which could be hidden in the real data. The artificial fully set the trade-off between the amount of exposures avail- campaigns also enable us to perform an arbitrary number of able to the optimization algorithm and the amount of expo- runs of the optimisation algorithms in order to evaluate their sures used for performance analysis. The more aggressive performance. This would not be possible on real campaigns, the optimization, the more biased (less reliable) are the es- since it is very expensive. timates of the variants’ performance. The balance of expo- sures across different variants must be such that the opti- To understand the reliability of the measured lift, it is crucial mization algorithm will yield the highest possible number to compute its confidence bounds. These can be computed of engagements, while still allowing estimation of the per- using the Fieller’s theorem [6] for the confidence interval of formance gain with a reasonable accuracy. Another obsta- the ratio of two means. To avoid the selection bias, we split cle, caused by technical limitations, is that feedback on how the exposures into two sets when running the campaign: the variants perform is usually delayed (in our case by a few control set, where the variants are displayed at random, and hours) and computed in batches (e.g., hourly). Therefore, the optimization set, where variants are displayed according the optimizing algorithm has to set the distribution of the to the decisions of the optimization algorithm. The control exposures across different variants on a hourly basis and not set is used to estimate the success rates of each variant and for each exposure separately – this has a negative effect on deduce the performance of the optimization algorithm. The the optimization potential of the algorithm. control set is usually small (∼ 10% of exposures), so the op- timization algorithm has a big enough volume of exposures 3. EVALUATING THE PERFORMANCE left to optimize. Low number of exposures in the control set has the disadvantage of confidence intervals being rather One approach for evaluating the performance of a digital ad- wide and the lower bound of the lift often being negative. We vertisement content optimization algorithm is to compute its improve the estimation of the bounds and still keep the same lift L = So − 1, where S S o is the number of successes achieved r 88 amount of exposures selected by the algorithm by introduc- pen more gradually a more conservative approach to for- ing three-set sampling [12]. Here, we split the optimization getting of historical data is needed. To deal with gradual set into a learning set and evaluation set. The control set changes in the trends we implement a two-memory struc- is composed of the exposures with randomly-selected vari- ture [9] of the historical data. We keep track of exposures ants. The learning set is composed of the exposures that and successes in two separate data structures: a long-term were selected by the optimization algorithm and are used (persistent) memory and short-term (transient) memory. The by it for further selection. The evaluation set is composed first holds almost all historical data and the latter holds only of the exposures that were selected by the algorithm, but very recent data about each variant. A weighted sum of the are not used by it for further selection – the optimization two is then used to estimate the success rates of the variants does not consider the exposures in the evaluation set when and feed them into the optimization algorithm. deciding which variant to chose next. Data from the con- trol set and the evaluation set can then be used to produce 5. DATA AND EXPERIMENTS unbiased estimates of the lift and its confidence bounds. We collected data from hundreds of advertising campaigns that used content optimization. The data includes the num- 4. OPTIMIZATION ALGORITHMS AND ber of exposures and successes for each variant for each hour THEIR IMPROVEMENTS of each campaign. Exposures in each campaign were split The optimization of advertisement variant selection can be in two groups, for 10% of the exposures a random adver- treated as the multi-armed bandit problem [10]. This is a tisement variant was chosen from an uniform distribution problem in which a gambler at a row of slot machines (one- (i.e., the control dataset), whereas for the remaining 90% armed bandits) has to decide which machines to play, how of the exposures a hand-tuned optimization algorithm chose many times to play each machine and in which order to play the variant. them. Each machine provides a random reward from a prob- ability distribution specific to that machine. The goal of the From all the campaigns, we selected a representative sample gambler is to maximize the sum of rewards earned through of 38 campaigns for further analysis. The durations of the a sequence of plays. A plethora of algorithms dedicated to selected campaigns span from 3 days to 6 months and the solving it exist [4], ranging from simple ones like the -greedy number of exposures ranges between 10, 000 and 10, 000, 000. selection policy, to more intricate ones like the Thompson There are 2 to 10 advertisement variants per campaign, with sampling (TS) [1], the upper confidence bound algorithm success rates ranging from 0.1% to 20% (median is 1.8%). (UCB) [2], and its improvements [3]. Each optimization The trend-change analysis shows there are 2 to 6 events per algorithm balances the exploration-exploitation trade-off us- campaign that change the success rates of the variants, apart ing its parameters. They instruct the algorithm what will from the monthly, weekly, and daily seasonalities. be the ratio between the number of exposures it will use to display the percievingly best variant (i.e., exploitation) and A preliminary analysis showed that 20 campaigns (out of 38) the number of exposures it will use to display the other vari- have at least some potential, so we used these to generate ants to see if some other variant is potentially better (i.e., two sets of 200 artificial campaigns: a best-case and worst- exploration). case set. We created 10 artificial campaigns from each real campaign (hence 200 from 20) per set. The two artificial sets By default, the aforementioned algorithms are not intended were then used to estimate the lower and upper bound of the to cope with delayed feedback, batch updates, and non- real campaigns’ potential and to measure the performance stationary data. Therefore, we propose several enhance- of the optimization algorithms. We tested three optimiza- ments to improve their performance on our use-case. A tion algorithms: ε-greedy, Thompson sampling, and UCB1- solution for the delayed feedback and batch updates is to Tuned. Each algorithm was run 10 times on each artificial simulate the immediate feedback for each variant selection campaign, producing 2000 runs per artificial set. Based on based on the historical success rates. When the real feed- these 2000 runs per algorithm we comparatively analysed back becomes available all of the simulated exposures are how many times each algorithm achieved lifts above 5%, discarded and replaced by the real data. 10%, and 20%. The simulations on artificial campaigns al- located 10% of exposures to the control set, 45% to the Non-stationarity of the data is problematic because the al- learning set, and 45% to the evaluation set. gorithms have trouble adapting to trend changes. We deal with abrupt changes in trends by using the Page-Hinkley 6. RESULTS AND DISCUSSION test [8] for detecting change-points in the success rate and We observe that approximately half of the campaigns have the number of exposure displayed per time unit. If the test no significant potential (Table 1): in the best case there is detects a change, we discard a portion of historical data to 45% of campaigns with lift below 5% and in the worst case enable faster adaptation of the optimization algorithm to there is 74% of such campaigns. A deeper analysis of the the new situation. Another approach that we propose is advertisement variants showed that there are considerable to perform periodic forgetting [10] of historical data based differences between the variants in some campaigns, whereas on some predefined condition like the total number of expo- the variants are nearly identical in other campaigns, which sures or time elapsed since the last forgetting event. After reduces the amount of optimization potential. the condition is met a forgetting event is triggered, which causes a part of the historical data to be discarded. In the initial experiments, we observed that ε-greedy is too aggressive for our non-stationary problem – it quickly fits The above two approaches are suitable if there are a lot of to a single variant and requires a long time to switch to abrupt changes in the trends. However, if the changes hap- another. Therefore, we omit it from further experiments 89 the potential of an advertising campaign reasonably well. Table 1: The theoretical potential in optimizing ad- However, many campaigns have no potential at all – cam- vertisement campaigns. paigns with no significant differences between variants do not benefit from optimization. In future work, we would Ratio [%] of real campaigns with potential lift above: like to identify what types of campaigns exhibit high poten- 5% 10% 20% tial. Discovering the characteristics that make a campaign suitable for optimization in the first place may significantly Worst case 26 18 8 increase the value of digital content optimization in digital Best case 55 53 34 advertising. Additionally, more advanced algorithms, like dual-layer UCB or budget-limited UCB [11], could be used to further increase the benefit of optimization. Another pos- Table 2: The performance of optimization algo- sible improvement would be to find the optimal relative sizes rithms. We present the lower and upper bounds of sets in the three-set sampling we used because large con- based on the results from the worst-case and best- trol and evaluation sets provide more data for the analysis, case sets of artificial campaigns, respectively. but inhibit the algorithms’ learning rate and hence perfor- mance. Ratio [%] of artificial campaigns with opt. lift above: 8. ACKNOWLEDGMENTS 5% 10% 20% We would like to thank Grega Kešpret and Gregor Smrekar Thompson sampl. 25–39 17–27 11–16 for the valuable insight they provided during our study. UCB1-Tuned 30–44 24–35 13–19 9. REFERENCES Imp. Thompson sampl. 36–55 23–37 12–30 [1] S. Agrawal and N. Goyal. Analysis of thompson Imp. UCB1-Tuned 34–59 27–35 14–30 sampling for the multi-armed bandit problem. In Conference on Learning Theory, pages 39–1, 2012. [2] P. Auer, N. Cesa-Bianchi, and P. Fischer. Finite-time and focus on the other two algorithms instead. Compar- Analysis of the Multiarmed Bandit Problem. Machine ison of the original Thompson sampling and UCB1-Tuned Learning, 47(2-3):235–256, 2002. algorithms shows the latter is better (Table 2). When the [3] P. Auer and R. Ortner. UCB revisited: Improved two algorithms are improved with the techniques presented regret bounds for the stochastic multi-armed bandit in Section 4, both Thompson sampling and UCB1-Tuned problem. Periodica Mathematica Hungarica, perform almost equally. The (improved) optimization algo- 61(1):55–65, 2010. rithms exploit approximately 25% to 50% of the potential [4] V. Kuleshov and D. Precup. Algorithms for in an advertisement campaign. Note that exploiting 100% multi-armed bandit problems. Journal of Machine of the potential is not possible. Learning, 1:1–32, 2014. We omit the in-depth analysis of the proposed improvements [5] T. Lu, D. Pál, and M. Pál. Contextual multi-armed of the optimization algorithms, and provide just a brief sum- bandits. In Proceedings of the Thirteenth international mary of the findings. We observed that simulated immedi- conference on Artificial Intelligence and Statistics, ate feedback only slightly improves performance, that two- pages 485–492, 2010. memory architecture is significantly beneficial, and that pe- [6] R. Moineddin, J. Beyene, and E. Boyle. On the riodic forgetting (with two memories) is as good as trend- location quotient confidence interval. Geographical change detection with the Page-Hinkley statistical test. We Analysis, 35(3):249–256, 2003. have not measured which forgetting approach is more resis- [7] S. Pandey and C. Olston. Handling advertisements of tant to parameter over-fitting. When the algorithms are im- unknown quality in search advertising. In Advances in proved with the techniques mentioned above, the exploratory neural information processing systems, pages parameters of the optimization algorithms can be adjusted 1065–1072, 2007. to make the algorithms more aggressive, since forgetting re- [8] R. Sebastiao and J. Gama. A study on change sets them frequently enough so they don’t stick to the same detection methods. In Progress in Artificial variant for too long. Intelligence, EPIA, pages 12–15, 2009. 7. CONCLUSIONS [9] D. Silver, R. S. Sutton, and M. Müller. Sample-based learning and search with permanent and transient In this study, we propose a methodology for estimating the memories. In Proceedings of the 25th international potential behind real-time optimization of digital advertise- conference on Machine learning, pages 968–975, 2008. ment content. We analysed hundreds of advertising cam- paigns provided by Celtra and developed a methodology for [10] R. S. Sutton and A. G. Barto. Reinforcement generating artificial campaigns that mimic real campaigns. Learning: An Introduction. The MIT Press, 1998. We generated multiple benchmark sets of artificial campaigns [11] L. Tran-Thanh, A. C. Chapman, A. Rogers, and N. R. and used them to empirically evaluate several optimization Jennings. Knapsack based optimal policies for algorithms in combination with different improvements. The budget-limited multi-armed bandits. In AAAI, 2012. proposed modifications proved highly beneficial and pro- [12] M. Xu, T. Qin, and T.-Y. Liu. Estimation bias in vided us with ideas that may increase the performance of multi-armed bandit algorithms for search advertising. the optimization algorithms even further. Our main dis- In Advances in Neural Information Processing covery is that optimization algorithms are able to exploit Systems, pages 2400–2408, 2013. 90 Indeks avtorjev / Author index Baert Anneleen ............................................................................................................................................................................. 67 Bizjak Jani .................................................................................................................................................................................... 27 Blatnik Robert .............................................................................................................................................................................. 19 Bohanec Marko ...................................................................................................................................................................... 63, 67 Boshkoska Biljana Mileva............................................................................................................................................................ 63 Bratko Ivan ................................................................................................................................................................................... 35 Brockhoff Dimo ........................................................................................................................................................................... 47 Budna Borut ................................................................................................................................................................................. 23 Ciancarelli Maria Costanza .......................................................................................................................................................... 67 Clays Els....................................................................................................................................................................................... 67 Cvetković Božidara ...................................................................................................................................................................... 75 Debeljak Marko ...................................................................................................................................................................... 51, 55 Dergan Tanja ................................................................................................................................................................................ 59 Dovgan Erik ........................................................................................................................................................................... 35, 67 Drobnič Vid .................................................................................................................................................................................. 75 Džeroski Sašo ................................................................................................................................................................... 11, 15, 51 Filipič Bogdan .............................................................................................................................................................................. 35 Fotiadis Dimitrios ......................................................................................................................................................................... 63 Gams Matjaž ................................................................................................................................................................ 7, 23, 39, 79 Gatsios Dimitris ........................................................................................................................................................................... 63 Gjoreski Martin ...................................................................................................................................................................... 23, 79 Gradišek Anton ...................................................................................................................................................................... 23, 27 Hansen Nikolaus .......................................................................................................................................................................... 47 Ilc Nejc ......................................................................................................................................................................................... 87 Janko Vito .................................................................................................................................................................................... 83 Konitsiotis Spiros ......................................................................................................................................................................... 63 Kononenko Igor ........................................................................................................................................................................... 31 Košir Domen ................................................................................................................................................................................ 87 Kostas M. Tsiouris ....................................................................................................................................................................... 63 Kuzmanovski Vladimir ................................................................................................................................................................ 55 Leprince Florence ......................................................................................................................................................................... 51 Luštrek Mitja .................................................................................................................................................. 43, 67, 71, 75, 79, 83 Marinko Matej .............................................................................................................................................................................. 71 Maslov Pavel ................................................................................................................................................................................ 67 Maver Aleš ................................................................................................................................................................................... 11 Miljković Dragana........................................................................................................................................................................ 63 Mitrevski Blagoj ........................................................................................................................................................................... 79 Mlakar Miha ................................................................................................................................................................................. 43 Novak Benjamin ........................................................................................................................................................................... 39 Pardaens Sofie .............................................................................................................................................................................. 67 Peev Gjorgi ................................................................................................................................................................................... 15 Petelin Gašper .............................................................................................................................................................................. 31 Peterlin Borut ............................................................................................................................................................................... 11 Petković Matej ............................................................................................................................................................................. 11 Piltaver Rok ............................................................................................................................................................................ 39, 87 Presetnik Primož .......................................................................................................................................................................... 27 Puddu Paolo Emilio ...................................................................................................................................................................... 67 Rigas George ................................................................................................................................................................................ 63 Schiariti Michele .......................................................................................................................................................................... 67 Šef Tomaž .................................................................................................................................................................................... 19 Simidjievski Nikola ...................................................................................................................................................................... 15 Slapničar Gašper .......................................................................................................................................................................... 71 Sluga Davor .................................................................................................................................................................................. 87 Sodnik Jaka .................................................................................................................................................................................. 35 Stensbye Andreas R. .................................................................................................................................................................... 83 Stepančič Luka ............................................................................................................................................................................. 27 91 Sušelj Gregor ................................................................................................................................................................................ 87 Tanevski Jovan ............................................................................................................................................................................. 11 Tosser Veronique ......................................................................................................................................................................... 55 Trajanov Aneta ....................................................................................................................................................................... 51, 55 Tušar Tea ...................................................................................................................................................................................... 47 Valmarska Anita ........................................................................................................................................................................... 63 Vidmar Lovro ............................................................................................................................................................................... 11 Vodopija Aljoša ........................................................................................................................................................................... 67 Vodopivec Tom ............................................................................................................................................................................ 87 92 Konferenca / Conference Uredili / Edited by Slovenska konferenca o umetni inteligenci / Slovenian Conference on Artificial Intelligence Matjaž Gams, Mitja Luštrek, Rok Piltaver Document Outline A - Naslovnica-SPREDNJA - A B - Naslovnica - notranja - A C- Kolofon - A D-E - IS2017 - skupni zacetni del Blank Page F - Kazalo - A G - Naslovnica podkonference - A H - Predgovor - A I - Programski odbor - A J - PDF - A 001 - Gams 002 - Petkovic 003 - Peev 004 - Sef 005 - Budna 006 - Bizjak 007 - Petelin 008 - Dovgan 009 - Novak Introduction MOLHC widget Hybrid-tree viewer MOLHC Evaluation Conclusion References 010 - Mlakar 011 - Tusar 012 - Debeljak-1 013 - Debeljak-2 014 - Dergan 015 - Boshkoska 016 - Bohanec 017 - Slapnicar 018 - Cvetkovic_IS2017 Introduction Related Work Materials and Methods Dataset Preprocessing Method Evaluation Conclusion ACKNOWLEDGMENTS References 019 - Gjoreski-NEW-NEW 020 - Janko 021 - Vodopivec K - Index - A L - Naslovnica-ZADNJA - A Blank Page Blank Page Blank Page Blank Page Blank Page