Informatica 38 (2014) 135-143 135 A Metaheuristic Approach for Propagation-Model TUning in LTE Networks Lucas Benedicic and Marko Pesko Research and Development Department Telekom Slovenije, d.d., Cigaletova 15, SI-1000 Ljubljana, Slovenia E-mail: {lucas.benedicic, marko.pesko}@telekom.si Tomaž Javornik Communication Systems Department Jožef Stefan Institute, Jamova cesta 39, SI-1000 Ljubljana, Slovenia Peter Korošec Computer Systems Department Jožef Stefan Institute, Jamova cesta 39, SI-1000 Ljubljana, Slovenia Keywords: LTE, coverage planning, semi-empirical propagation models, metaheuristics, optimization Received: November 28, 2013 When deploying a new mobile technology such as LTE, it is crucial to identify the factors that affect the radio network in terms of capacity and quality of service. In this context, network coverage is arguably the single most influential factor. This work presents a metaheuristic-optimization approach that automatically adapts the parameters of a signal-propagation model. The optimization procedure is performed per cell, enabling the calculation of accurate network-coverage predictions. The evaluation of the proposed approach is carried out on two different regions in Slovenia, where Telekom Slovenije, d.d., provides LTE coverage. The results show radio-propagation predictions of improved quality when compared to manual and analytical methods. Povzetek: Avtorji predstavljajo metahevristično metodo za samodejno optimizacijo parametrov semi-empiricnih modelov razširjanja radijskega valovanja. Rezultati številnih poskusov v omrežjih LTE kažejo izboljšano natančnost izračunanih napovedi razširjanja radijskega valovanja. 1 Introduction One of the primary objectives of radio-coverage planning is to efficiently use the allocated frequency band. To this end, radio-coverage prediction tools are of great importance, as they allow network engineers to test different configurations before physically implementing the changes. However, predicting the radio coverage of a mobile network is a complex task, hence the importance of fast and accurate prediction tools. The precision achieved by the software tool of choice is directly related to the accuracy of the signal-propagation model used. For this reason, signal-propagation models that support configurable parameters are preferred, since they allow the model to adapt to different environments and thus to improve the accuracy of the calculated coverage predictions. The effectiveness of the decision-making process during radio-network planning is tightly coupled with the precision achieved by the propagation model used. In order to obtain a radio-propagation model that accurately reflects the characteristics of the area covered by the mobile network, the parameters of the signal-propagation model are adjusted using data from field-measurement campaigns. Signal-loss adjustment using this method depends on existing field-measurement data, which are collected in advance for the area covered by the target network cells. In order to adapt the parameters of a signal-propagation model, mainly analytical approaches were proposed in the related literature [1, 8, 27]. These works confirm the suitability of methods based on least-squares theory for the parameter tuning of signal-propagation models. In this work, we propose the automatic optimization of the parameters of a signal-propagation model using a metaheuristic approach. The objective of such optimization is four-fold. First, using a stochastic optimization approach, the parameters are optimized in order to reflect the local characteristics of the terrain, thus adapting the model to the local environment of a given network cell. Second, based on a set of field measurements, the automatic optimization of model parameters improves the accuracy of the calculated radio-propagation predictions of each network cell, as well as the radio network as a whole. Third, the proposed metaheuristic method improves, under certain geographical conditions, the results achieved by a traditional linear-least-squares approach [1, 8, 27], the application of which only adapts the linear part of the propagation model, i.e., y = c + mx. Fourth, tuning the complete parameter set per network cell shows improved results especially in rugged-terrain areas. As the working schema for tackling the presented prob- 136 Informatica 38 (2014) 135-143 L. Benedicic et al. lem, we use PRATO, a parallel framework for coverage planning of cellular networks [5]. The framework flexibility allows for coverage planning and optimization of radio networks in general, and LTE in particular. The optimization component featured by PRATO enables the parallel optimization of several parameters, e.g., the parameters of the signal-propagation model. The remaining of this paper is organized as follows. Section 2 introduces some principles of radio-propagation prediction and the signal-propagation model used. Section 3 describes the optimization problem involving the parameter-tuning of the signal-propagation model, followed by the performed simulations and the achieved results in Section 4. Finally, in Section 5 we draw some conclusions and give guidelines for further work. 2 Radio-propagation prediction To calculate the radio-propagation predictions, we use a mathematical model based on the well-known Okumura-Hata formula [14, 22]. Other more accurate methods exist, like the ones based on ray tracing [7, 25]. However, these methods are still inefficient in terms of the computational effort required to achieve satisfying results. On the other hand, (semi-)empirical methods for radio-propagation predictions give acceptable results within a feasible amount of time. For this reason, they became the industry standard for non-deterministic, signal-propagation calculations [3, 6, 14, 21, 22, 23]. 2.1 Signal-propagation model The Okumura-Hata model has been largely studied and shown to be suitable for predicting the signal propagation of LTE networks [2]. In its primary form, the model distinguishes the distance from the receiver to the transmitter, the frequency used and the effective antenna height, i.e., the antenna height above the receiver's level. These variables are taken into account in order to calculate the path loss in open areas (OA), as described in Equation (1). LoA(x,y, £) = £1 + £2 log(d(x,y)) + £3 log(HA) + £4 log(d(x,y)) log(HA) - 3.2 (log(11.75 • Hr))2 + 44.49 log(F) - 4.78 (log(F))2 , (1) where 3 = (¡1, ¡2, ¡3, ¡4) are the adaptable parameters of the model, d(xy) is the distance from the transmitter to the topography point with coordinates (x, y) (expressed in kilometers), HA is the effective antenna height of the transmitter (expressed in meters), HR is the antenna height of the receiver (expressed in meters), and F is the frequency (expressed in MHz). In this work, as well as in [11], the terrain profile is used for non-line-of-sight (NLOS) determination, i.e., the loss due to an obstacle obstruction in the first Fresnel zone of the transmitter [26]. In such case, additional path-loss factors due to the terrain profile and the Earth shape are added to the original formula, the values of which are calculated as in Equation (2). Lnlos(x, y) = yj{aK(d(Xy)))2 + E(d(Xy))2, (2) where a is the knife-edge diffraction control parameter, the value of which is calculated based on the level of obstruction of the Fresnel zone, K(d(x,y)) is the knife-edge diffraction loss (in dB), and E(d(x,y)) is the correction due to the Earth sphere, the value of which improves the calculated prediction especially for higher base-station towers and distances over 10 kilometers. All three values depend on the characteristics of the topography point with coordinates (x, y). The euclidean distance between the transmitter and the receiver is intentionally calculated using two-dimensional coordinates due to simplicity and the negligible difference when compared to its three-dimensional counterpart. In order to adequately predict signal-loss effects due to foliage, buildings and other fabricated structures, a supplementary factor based on the land usage (clutter) is included. This technique is adopted by several propagation models for radio networks, e.g., [1, 3, 19, 22]. Consequently, we introduce an extra term for signal loss due to clutter, thus defining the total model-predicted path loss, which is expressed in dB, as in Equation (3). L(x, y, £) = Loa(x, y, £) + Lnlos(x, y) + Lclut(x, y), (3) where LCLUT(x, y) represents the clutter loss at the topography point with coordinates (x, y). 2.2 Field measurements In mobile networks, a moving mobile device (or user equipment, UE) constantly performs cell selec-tion/reselection and handover in order to keep the best possible connection to the network. In this context, the best connection is selected by measuring the signal strength or quality of the neighboring cells. In LTE networks, the UE measures two parameters from the reference signal of the network, namely the Reference Signal Received Power (RSRP) and the Reference Signal Received Quality (RSRQ) [23]. For a certain frequency bandwidth, RSRP measures the average received power over the resource elements that carry cell-specific reference signals. RSRP is applicable in both idle (e.g., waiting for a call) and connected (e.g., during a call) modes. During the procedure of cell selec-tion/reselection in idle mode, RSRP is used. On the other hand, RSRQ is only applicable when the UE is in connected mode. The radio-coverage calculation involves predicting the network coverage over a certain region, and thus over the UEs within it. Hence, in the first place, we are interested on accurately predicting the best connection the UE would select in idle mode and the RSRP field measurements it uses. A Metaheuristic Approach for Propagation-Model Tuning in LTE Networks Informatica 38 (2014) 135-143 137 In our case, the field measurements representing the RSRP at a given location were collected using a small truck equipped with the spectrum analyzer Rohde & Schwarz, the functionality of which supports LTE signal analysis. The spectrum analyzer was connected to an external omnidirectional antenna mounted on the roof of the truck, at roughly 2 m above the ground, taking measurements at a rate of 2 Hz. The measurement locations were established using a GPS unit. These GPS-informed locations were tested to be compliant with the 60-meter limit mentioned in [1]. The measurements covered a considerable proportion of the target area, with over 100,000 individual points, collected from more than 30 network cells. To minimize the deviation in the measured RSRP values, and the impact that small-scale fading has in larger-scale path loss [10], all field measurements were post-processed so that a single value, the median, was calculated for each of the measured locations. The resulting RSRP was then used as the field measurement representing the given location, the resolution of which matches the digital elevation model (DEM) and clutter maps used as input data of the optimization process. Note that the DEM data represents the terrain profile of the geographical area of interest. 3 Parameter tuning of the radio-prediction model The procedure to adapt the parameters of the mathematical model for each of the cells in the target network involves minimizing the deviation of the radio-propagation prediction compared to a given set of field measurements. In the context of our work, this means that we have to fine tune the parameters of LOA(x, y, ¡3), as defined in Equation (3). The adjustable parameters are the elements of vector 3 = (¡1, ¡2, ¡3, ¡4) € R4, namely ¡1 the reference loss or offset; ¡2 the loss slope due to distance of the receiver from the transmitter; ¡3 the loss slope due to height of the transmitter antenna; ¡4 the loss slope due to the combined effect of the distance and height of the antenna. The parameter tuning is performed per cell to improve local fitting of the radio predictions, being its resulting solution a vector I3* of the target cell. The analytical approach for tuning of the radio-prediction model consists of correlating the field measurements with the predicted received-signal values. The new parameter set originates from the minimization of an error criterion. As defined in [15, 27], the minimization criterion is the squared-sum difference between the predicted and the observed RSRP levels, the definition of which is shown later in Equation (6). As a general rule when applying this approach, only the first two components of the vector ¡3 are adapted, i.e., ¡1 and ¡2, whereas the values of ¡3 and ¡4 are kept constant. Therefore, the analytical method consists in fitting only the linear part of the path-loss definition previously presented in Equation (3), i.e.: AL(x,yJ) = 31 + ¡2 log(d(x,y)). (4) The expression in Equation (4) does not take the terrain height into account, which is feasible when the field measurements are taken at a roughly constant height relative to the base station [8, 27]. However, when these heights fluctuate within the coverage area of a cell, the other two parameters, ¡3 and ¡4, have a considerable effect on the adaptation of the signal-propagation model, as it will be shown in the following sections. 3.1 Differential ant-stigmergy algorithm In order to adapt the vector 3, including its four components, we turn our attention to metaheuristic algorithms in general [24] and swarm intelligence in particular [16]. From this last family of metaheuristics, we have chosen the differential ant-stigmergy algorithm (DASA) [17]. A standalone metaheuristic, the DASA is based on the well-known Ant-Colony Optimization (ACO) [9]. It provides a specialized extension for solving high-dimensional, numerical-optimization problems, whereas the ACO operates on the discrete domain. The DASA represents the search space in a fine-grained discrete form, producing a graph. This graph is then used as the walking paths for the ants, which iteratively improve the temporary best solution. There are several reasons for choosing the DASA as the optimization algorithm in the context of this problem. First, the benefits of metaheuristic algorithms for solving optimization problems, particularly in the context of radio networks, was demonstrated by numerous authors [4, 12,15,18]. Second, in [17], the authors shown the suitability of the algorithm for solving numerical problems, also exhibiting better performance than other swarm-based metaheuristics. Moreover, it has already been successfully applied for tackling an optimization problem in the area of radio networks [4], obtaining competitive results. The mapping between the parameter-optimization problem and the DASA is as defined in Equation (5). Xa = {xi, X2, X3, X4} , (5) where Xa is the solution vector of ant a during the minimization process, and xj represents the j-th component of vector ¡3 for the signal-propagation model of a given cell. At the end of every iteration, and after all the ants have created solutions, they are evaluated to establish if any of them is better than the best solution found so far. For a more in-depth explanation about this procedure and the DASA itself, we refer the reader to [17]. 138 Informatica 38 (2014) 135-143 L. Benedicic et al. Network-cell configuration Network cell Drive tests ^ Ceil field i Optimization measurements Radio-propagation prediction i Stopping criteria Optimized cell a parameters Figure 1: The parameter-optimization system as executed per network cell by a parallel process of PRATO. 3.2 Optimization objective The optimization objective consists in adjusting the values of components of vector 3 according to a set of field measurements of a given cell. Each network cell is independently optimized, so that its radio-propagation prediction minimizes the mean-squared error against the field measurements, as defined in Equation (6). f*(c) = min V (pc - Lc(c°°rd(m)'^) - m)2 Vc € N, (6) f ( ) JtMc Mcl ' ( ) where f *(c) is the optimization objective to be minimized for cell c, N is one of the test networks, pc is the transmit power of cell c, m is a field measurement of cell c, Mc is the set of all field measurements of cell c, and Lc(coord(m), 3) represents the path loss of cell c at the same geographical point of the field measurement m, as defined in Equation (3). Note that f *(c) is independently calculated for each c N. 4 Simulations In the parameter-optimization problem, the process starts with a mobile network. Each network cell is optimized by an independent parallel process as shown in Figure 1. The set of field measurements corresponding to the cell under optimization has to be gathered through drive tests before hand. Together with the cell configuration, they provide the input data for the optimization process itself. An iteration begins when the DASA generates a solution vector for each of the ants in the colony. The following step involves the evaluation of the solution vector carried by an ant, i.e., one radio-propagation prediction per iteration of the optimization process. The objective-function value is calculated as defined in Equation (6), and sent back to the DASA for it to generate the next set of solutions. The optimization process involves multiple iterations, which are repeated until some stopping criteria are met. Then, the best solution found represents the optimized values of the tuning parameters for the radio-propagation model of the target network cell. The optimization process is performed by PRATO in parallel over the worker processes, each of which runs independently of the others while optimizing the parameters of one network cell. Compared with the analytical approach, the solution of which requires solving a linear system of equations, a large number of evaluations is needed for the metaheuristic optimization to converge to a solution. Therefore, it is essential Number of cells Calculation radius [km] Area [km2] Field-measurement proportion [%] Neti Net2 9 25 16.00 16.00 82.90 133.47 4.40 6.74 Table 1: Some characteristics of the test networks used for the experimental simulations. to exploit the parallel nature of PRATO in order to concurrently optimize the parameter sets of multiple cells within the network. Otherwise, such approach would not be feasible, since the time required to reach a reasonable solution would be excessive. 4.1 Test networks The test networks, Neti and Net2, are subsets of a real LTE network deployed in Slovenia by Telekom Slovenije, d.d. For the path-loss predictions, we were provided DEM and clutter maps of 25 m2 resolution. A calculation radius around each network cell limited the path-loss prediction to a distance where it is feasible for an UE to connect to a cell, i.e., when the RSRP is greater or equal to -124 dBm [20]. At the same time, this calculation radius provides enough overlap among neighboring cells to calculate the network coverage over the whole region, for which the receiver height was set to 2 m above ground level. Table 1 provides more information about the test networks used, such as the number of network cells, the area surface, and the covering proportion of the collected field measurements in terms of the total area of each test network. Neti represents a network deployed over a dominant agricultural area with almost flat terrain, some forests and waters streams. The other network, Net2, is deployed over a hilly terrain mostly covered by forests. As the stopping criteria for the optimization runs, we fixed the maximum number of iterations to 250, since the algorithm showed an acceptable convergence profile in all runs. Overall, the framework completed 20,000 objective-function evaluations, i.e., 180,000 radio-coverage predictions for Neti and 500,000 for Net2. The simulations were carried out on several computing nodes of the DEGIMA cluster [13] at the Nagasaki Advanced Computing Center (NACC) of the Nagasaki University in Japan. The reason for using a high-end computer cluster as DEGIMA is to exploit the parallel nature of PRATO. To this end, groups of 5 and 13 computing A Metaheuristic Approach for Propagation-Model Tuning in LTE Networks Informatica 38 (2014) 135-143 139 Manual Analytical DASA Test network Mean Standard deviation Mean Standard deviation Mean Standard deviation Net1 5.88496 13.69240 0.00001 12.64708 0.01974 12.13723 Net2 6.52300 14.35561 0.00007 12.30833 0.01372 10.59590 Table 2: Mean and standard-deviation values of the radio-propagation prediction against the field measurements. The values, expressed in dB, are given when using the manual, analytical and DASA approaches in each test network. nodes were used for executing the simulations of the different problem instances, i.e., Neti and Net2, respectively. The computing nodes were connected by a LAN, over a Gigabit Ethernet interconnect. The nodes were equipped with a Linux 64-bit operating system (Fedora distribution). OpenMPI was used as the message passing implementation, version 1.6.1, the binaries of which were manually compiled with the distribution-supplied gcc compiler, version 4.4.4. After some trial-optimization runs, the six parameters that control the way the DASA explores the search space were set to the following values: - m = 80, the number of ants; - b =10, the discrete base; - p = 0.2, the pheromone dispersion factor; - s+ = 0.01, the global scale-increasing factor; - s— = 0.01, the global scale-decreasing factor; and - e = 10—5, the maximum parameter precision. The trial runs consisted in doubling m from 5 to 640, and verifying the convergence profile and best solution found. The values of the other parameters were left unchanged. 4.2 Result analysis The mean and standard-deviation values after correlating the field measurements with the radio-propagation prediction of each test network and parameter set are shown in Table 2. The parameter set used for the "Manual" column of Table 2 was provided by the radio experts of the Radio Network Department at Telekom Slovenije, d.d. These values were calculated based on manual observations and were applied to all the cells in the network. The column labeled as "Analytical" represents the parameter set calculated by the least-squares approach as presented in the related literature [1, 8, 27]. The last column, which is labeled as "DASA", represents the average parameter set calculated by the DASA after 30 independent runs. The manual approach clearly shows the biggest discrepancy of the radio-propagation predictions in both the rural (Net1) and hilly (Net2) environments. The analytical approach considerably improves the results achieved by the manual method, thus showing lower mean and standarddeviation values in both test networks. As for the DASA, it further improved the standard deviation of the analytical approach. Moreover, this correction is significant for in Net2, thus confirming the influence of the hilly terrain in the accuracy of signal-propagation predictions. This is especially important on the border of the cell coverage, where a 2 dB difference in the received-signal strength could mean predicting sufficient network coverage where there would otherwise be none. Regarding the mean values showed by the DASA solutions, we may observe that they are several orders of magnitude higher than those of the analytical solutions. However, the values are, in all cases, strictly lower than 0.02 dB, which is a negligible difference in terms of the RSRP levels that outline the coverage of a network cell. These results confirm that the use of the DASA to perform the optimization of parameters of a signal-propagation model is viable, since it is capable of reflecting the physical phenomena appearing in real-world conditions in two geographically-different network instances. Figures 2 and 3 depict the probability-density distributions of the difference between the signal-propagation predictions and the field measurements. The mean and standard-deviation values of these distributions are listed in Table 2. Figure 2 (a) depicts the difference distribution of the coverage prediction for test network Neti using the manually-calculated parameters, Figure 2 (b) shows the difference distribution for the same test network, but using the analytically-calculated parameters, and Figure 2 (c) shows the difference distribution using the DASA-optimized parameters. Notice how the difference distributions show an improvement when the analytically-calculated parameters are used, lowering the largest (outer) deviations, and raising the lowest (inner) ones. Additionally, the difference is negligible when using the optimized parameters, thus confirming that it is sufficient to only adapt ¡0 and ¡1 in environments where the height difference between the transmitter and the receiver, i.e., log(HA) in Equation (1)), is roughly constant. The difference distributions of the radio-propagation predictions for test network Net2 using the manually-calculated parameters, the analytically-calculated, and the optimized ones are shown in Figures 3 (a), 3 (b), and 3 (c), respectively. Similar to Net1, the improvement appears in the largest deviations, since their values are lower than when using the manually-calculated parameters. In this case, we may also observe the improvement achieved by the parameter set calculated with the DASA, which included all four components of the vector ¡3, thus better re- 140 Informatica 38 (2014) 135-143 L. Benedicic et al. 0.045 |-r 0.040 — 0.035 — 0.030 -0.025 -0.020 -0.015 -0.010 -0.005 0.000 > ': : o 7o if 4> ^ 0.045 0.040 0.035 0.030 0.025 0.020 0.015 0.010 0.005 0.000 ■ ■':: :: ^ if h. J I'VVW V 0.045 pr 0.040 — 0.035 — 0.030 -0.025 -0.020 -0.015 -0.010 -0.005 0.000 Difference [dB] (a) <%. <% <% <% <% <%„ .<% <%. -'Oj 0.045 0.000 Difference [dB] (a) Oy C^i -Oy "C^i 'Oj 'Oj \"% "'Oj "'Oj"°J ® ® ® ^ Difference [dB] Difference [dB] 0.045 pr 0.040 — 0.035 -0.030 -0.025 -0.020 -0.015 -0.010 -0.005 0.000 (b) \"% "'«<® °j ® ® ' 0.045 0.040 0.035 0.030 0.025 0.020 0.015 0.010 0.005 0.000 (b)