42517 KB0 KB0 KB2019Džeroski, SašoJevšenak, JernejLevanič, TomXII, 141 f., 31 cmCOBISSID:930167PID:https://repozitorij.uni-lj.si/IzpisGradiva.php?id=105984URN:URN:NBN:SI:doc-RY8KSEF1enslJ. Jevšenakvisokošolska delabranikedendrokronologijadisertacijelinearna regresijametodemodelna drevesanelinearne metodeokoljske spremenljivkeparametriregresijska drevesastrojno učenjeumetne nevronske mrežeNelinearno modeliranje povezav med lesnimi branikami in okoljem| doctoral dissertation| doktorska disertacija| Nonlinear modeling of the relationship between xylem tree-rings and environment|(Multiple) linear regression (MLR) and selected nonlinear methods from the field of machine learning were compared for the analysis of relationships between xylem tree-rings and the environment: artificial neural networks with a training algorithm that uses Bayesian regularization (ANN), model trees (MT), ensembles of model trees (BMT) and random forests of regression trees (RF). The selected methods were compared on nine datasets, which included different tree-ring parameters and different target environmental variables. For the nonlinear methods, better statistical metrics were calculated on validation data in most cases, but the differences in comparison to linear regression were minor. Additional analysis indicated that the methods mostly differ in predicting the extreme values. The characteristic of nonlinear methods is that the change in the dependent variable is not proportional to the change of one or more independent variables. The latter results in a reduced range and variability of reconstructed values, which makes the reconstruction visually less attractive as compared to linear extrapolation, even though in most cases statistically better. None of the nonlinear machine learning methods showed best results on all datasets, therefore it makes sense to always compare different machine learning regression methods prior to climate reconstruction. To do so, the R function compare_methods() was developed and implemented in the dendroTools R package, which is freely available on the CRAN repositoryZa analizo povezav med lesnimi branikami in okoljem so primerjane (multipla) linearna regresija (MLR) in izbrane nelinearne metode s področja strojnega učenja: umetne nevronske mreže z učnim algoritmom Bayesova regularizacija (ANN), modelna drevesa (MT), ansambli modelnih dreves (BMT) in naključni gozdovi regresijskih dreves (RF). Izbrane metode so bile primerjane na devetih podatkovnih množicah, ki so vključevale različne parametre branik in različne okoljske spremenljivke. Za nelinearne metode so bili v večini primerov izračunani boljši statistični kazalci za validacijsko množico, čeprav razlike v primerjavi z linearno regresijo niso bile velike. Dodatne analize so pokazale, da se metode večinoma razlikujejo pri napovedih ekstremnih vrednosti. Značilnost nelinearnih metod je, da sprememba odvisne spremenljivke ni premo sorazmerna spremembi ene ali več neodvisnih spremenljivk. Slednje pomeni zmanjšan razpon in variabilnost rekonstruiranih vrednosti, kar rekonstrukcijo naredi vizualno manj privlačno od linearne ekstrapolacije, čeprav v večini primerov statistično boljšo. Nobena izmed nelinearnih metod s področja strojnega učenja ni dala najboljših rezultatov na vseh podatkovnih množicah, zato je pred rekonstrukcijo klime z metodami strojnega učenja vedno smiselno primerjati več metod. Za ta namen smo razvili R funkcijo compare_methods() in jo vgradili v dendroTools R paket, ki je prosto dostopen na CRAN-repozitorijuTEXTvisokošolska delatheses and dissertationsSlovenian National E-content AggregatorNational and University Library of SloveniaUniverza v Ljubljani, Biotehniška fakulteta