A study on machine learning and regression based models for performance estimation of LTE HetNets B. Bojović 1, E. Meshkova 2, N. Baldo 1, J. Riihijärvi 2 and M. Petrova 2 1 Centre Tecnològic de Telecomunicacions de Catalunya (CTTC) Castelldefels, Spain 2 Institute for Networked Systems, RWTH Aachen University, Aachen, Germany ACROPOLIS 3rd Annual Workshop - London, 2013
Outline Motivation 1 Motivation Femtocell network deployments Inter-cell interference awareness Dynamic frequency allocation for LTE system 2 LTE physical layer Proposed prediction methods and covariates 3 4 Machine learning techniques for performance estimation
Femtocell network deployments Goals and issues. Femtocell network deployments Inter-cell interference awareness Dynamic frequency allocation for LTE system Recently, there was much interest in femtocells deployments for different network technologies. The most of attention was in shown for WCDMA technology in 3G network deployments The same concept could be applied also to more recent technologies such as WiMAX, LTE,.. The goal is to improve the performance of the mobile network by: extending indoor coverage achieving higher data rate for the indoor and short range outdoor communications The main challenge is how to manage the available spectrum resources
Femtocell network deployments. Goals and issues. Femtocell network deployments Inter-cell interference awareness Dynamic frequency allocation for LTE system One approach could be to divide the available spectrum into several frequency bands and then to assign to each femtocell a different one In previous case, the drawback are: low or no frequency reuse, not feasible for dense deployments and requires much maintenance Another approach, which is typically considered in recent technologies, is to share available frequency bands among femtocells In the previous case the main issue is the interference between femtocells, between femtocell and macro cells Consequence: the network performance degradation and unfairness among users
Inter-cell interference awareness Previous work Femtocell network deployments Inter-cell interference awareness Dynamic frequency allocation for LTE system Many studies have identified significant performance gains in the systems that are aware of the inter-cell interference Objective: to maximize the system performance and to provide fairness among users by minimizing the inter-cell interference Some of proposed techniques are based on: power control schemes, static and adaptive fractional frequency reuse schemes, cancellation techniques, intelligent scheduling, network power coordination,...
Femtocell network deployments Inter-cell interference awareness Dynamic frequency allocation for LTE system Dynamic frequency allocation for LTE system Our approach Our approach it to improve the system performance by enabling the dynamic frequency allocation We consider an LTE system, which is designed to be used with frequency reuse factor of 1 LTE technology includes advanced features such as OFDMA, SC-FDMA, AMC, dynamic MAC scheduling and HARQ,.. More difficult than in previous technologies to predict the system capacity for specified system configuration Our approach is to apply machine learning and regression analysis for system capacity estimation, that will enable efficient dynamic frequency allocation
LTE physical layer Configurable network parameters LTE physical layer Proposed prediction methods and covariates The OFDMA multicarrier technology provides flexible multiple-access scheme that allows different spectrum bandwidths to be utilized without changing the fundamental system parameters or equipment design In this work we consider: 1 the carrier frequency f c ; and 2 he bandwidth B In frequency domain resources are grouped in units of 12 subcarriers that are called resource blocks centered around f c and occupy 180 khz in frequency domain
LTE physical layer Proposed prediction methods and covariates Configuring carrier frequency and bandwidth. An example The channel raster is 100 khz for all bands LTE enb operates using a set of B contiguous resource blocks(rb); the allowed values for B are 6, 15, 25, 50, 75, 100 RBs ( 1.4, 3, 5, 10, 15, 20 MHz) Femtocell 1 f o1 EARFCN f c1 f b1 180kHz 600kHz Femtocell 2 f o2 EARFCN f c2 f b2 f b1 - f o2
LTE physical layer Proposed prediction methods and covariates The optimization problem in broader context The quality of the achieved performance depends on various factors: amount of available inputs, their usability in the context of particular models and optimization methods Inputs Models & Decisions Propagation losses, MAC and application layer throughput 2-4 node topology Parameters & KPIs Schedulers Topology Time granularity Spatial sampling Per user/femto reporting System information/dependencies Performance Abstractions/simplifications Methods Online adaptation and/or training Required parameteres and prior info General criteria Criteria specifically discussed in our work Coloring, Graph abstraction, SINR, and MAC throughput estimation Regression analysis (several forms) Genetic algorithms Derived system dependencies (PHY - MAC(incl. schedulers) - Application) Complexity Training time Robustness Reusability Accuracy Optimality Num. of samples for training Prediction accuracy Comments on computation effords Per network/femto/user predictions
The specific optimization problem An example LTE physical layer Proposed prediction methods and covariates The specific problem that we are solving is: select for each for each deployed enb i = 1,..., N, the frequency f c and the system bandwidth B that will provide the best network performance. In this work we consider capacity as performance metric (we are interested to consider delay, fairness, different QoS metrics) Another aspect to consider is the number of possible solutions, which is exponential with N An example for 4 femtocell scenario, total available bandwidth of 5 MHz, f c muliple of 300 khz and B that can take value 6,15,25 there are 4625 physically distinct solutions
LTE femtocell capacity estimation design the cost function LTE physical layer Proposed prediction methods and covariates Common approach is to use some variation of Shannon formula, but those approaches typically do not consider effects of AMC, MAC scheduling,.. The TB size for each given AMC scheme and number of RBs is defined by the LTE specification We consider effect of different schedulers and for that purpose we use Round Robin and Proportional Fair Capacity in bits/rb 1500 1000 500 LTE Shannon mod.shannon 0 10 5 0 5 10 15 20 25 30 SINR
prediction methods and covariates LTE physical layer Proposed prediction methods and covariates Our goal is to predict the performance of the femtocell accurately by using regression analysis and machine learning techniques The information used: basic pathloss and configuration information, a limited number of feedback measurements that provide the throughput and the delay metrics for a particular frequency settings Different regressors: SINR, SINR/MAC throughput mapping and different sampling technique affects the prediction performance The performance metrics considered are: network-wide and per-user throughput The methods considered: Bagging tree, Boosted tree, Kohonen network, SVM radial, K-nearest neighbor, Projection Bojović, Meshkova, pursuit Baldo, Riihijärvi, regression, Petrova Machine Linear Learning and regression for performance estimation
. To simulate this scenario we use LENA, LTE-EPC network simulator We choose a typical LTE urban scenario with buildings (in LTE literature known as dual stripe scenario) The HeNbs are randomly distributed in the buildings and each HeNb has equal number of users For the four femtocell scenarios we simulate: 1 three users per node allocation with the total of 2 MHz bandwidth, and 2 two users per node allocation with the total of 5 MHz bandwidth
Dual-stripe scenario Radio environmental map in a scenario with 2 buildings and 2 HeNbs 300 20 y-coordinate [m] 280 260 240 220 200 220 240 260 280 300 320 340 x-coordinate [m] 15 10 5 0-5 SINR [db]
Scenario 1 Machine learning techniques for performance estimation Scenario with 4 femtocells and 3 users each femto, using 2 MHz overall bandwidth, showing actual measured MAC throughput vs. Shannon models. Sum SINR means a sum over all RBs Even if correlations between the different SINR related metrics and the achieved throughputs are apparent, the dispersion of the results is rather substantial
Scenario 1 Machine learning techniques for performance estimation MAC Throughput [Mbps] 6.0 7.0 8.0 9.0 1000 1200 1400 1600 6.5 8.0 6.0 7.0 7.5 8.5 Sum SINR [db] Sum SINR/THR NS3 Mapping [Mbps] (a) Measured MAC throughput vs. min SINR (left) and SINR/MAC thoughput mapping (right) for a scenario with 4 femtos with 3 users each, Proportional Fair scheduler, and TCP traffic. MAC Throughput [kbps] 600 800 1000 1200 1400 MAC Throughput [kbps] 100 150 200 250 300 MAC Throughput [Mbps] 5.5 6.0 6.5 7.0 7.5 MAC Throughput [Mbps] 5.5 6.0 6.5 7.0 7.5 MAC Throughput [Mbps] 6.0 7.0 8.0 9.0 1000 1200 1400 1600 6.5 8.0 6.0 7.0 7.5 8.5 Sum SINR [db] Sum SINR/THR NS3 Mapping [Mbps] (b) Measured MAC throughput vs. min SINR (left) and SINR/MAC thoughput mapping (right) for a scenario with 4 femtos with 3 users each, Round Robin scheduler, and TCP traffic. 600 720 840 960 1080 1200 120 240 360 480 Sum SINR/THR NS3 Mapping [kbps] Sum SINR/THR NS3 Mapping [kbps] (c) Measured MAC thoughput for two selected users vs. SINR/MAC THR mapping for a scenario with 4 femtos 3 users each, Proportional Fair (left) and Round Robin schedulers (right) for TCP traffic.
Scenario 2 Machine learning techniques for performance estimation Scenarios with 2 femtocells and 5 users each and 4 femtocells with 2 users each, using a bandwidth of 5 MHz. Different behavior is noted for UDP and TCP traffic, and for different schedulers From these results we can expect effective performance prediction to be a challenging problem especially for TCP traffic and a small number of users per femtocell.
Scenario 2 Machine learning techniques for performance estimation 4.0 MAC Throughput [Mbps] 6.0 8.0 10.0 4.0 MAC Throughput [Mpbs] 6.0 8.0 10.0 4.0 8.0 12.0 16.0 20.0 0 5 10 15 20 Min SINR per RB [db] Sum SINR to MAC Throughput Mapping [Mbps] (a) Measured MAC throughput vs. min SINR (left) and SINR/MAC thoughput mapping (right) for a scenario with 4 femtos with 2 users each, Proportional Fair scheduler, and TCP traffic. MAC Throughput [Mpbs] 10.0 15.0 20.0 25.0 MAC Throughput [Mpbs] 10.0 15.0 20.0 25.0 0 5 10 15 20 Min SINR per RB [db] 4.0 8.0 12.0 16.0 20.0 Sum SINR to MAC Throughput Mapping [Mbps] (b) Measured MAC throughput vs. min SINR (left) and SINR/MAC thoughput mapping (right) for a scenario with 4 femtos with 2 users each, Proportional Fair scheduler, and UDP traffic. MAC Throughput [Mpbs] 2.0 3.0 4.0 5.0 MAC Throughput [Mpbs] 4.0 6.0 8.0 10.0 0.8 1.2 1.6 2.0 2.4 Sum SINR to MAC Throughput Mapping [Mbps] 2.0 3.0 4.0 5.0 Sum SINR to MAC Throughput Mapping [Mbps] (c) Measured MAC throughput vs. SINR/MAC thoughput mapping for a selected femtocell (left) and a whole network (right) for a scenario with 2 femtos with 10 users each, Round Robin scheduler, and UDP traffic.
Pursuit regression technique Machine learning techniques for performance estimation Scenario with 4 femtocell and 3 users per each femto and bandwidth of 2MHz Random sampling with 10% of 337 permutations being explored The regression technique is applied The earlier expectation that UDP traffic with more primitive scheduler results in higher predictability is confirmed, the users the predictions are very accurate
Pursuit regression technique Figure Machine learning techniques for performance estimation Throughput [Mbps] 0 0.5 1.0 1.5 Measured Predicted 0 50 100 150 200 250 300 350 Permutation (a) Measured and predicted MAC throughput for a selected user with permutations ordered after measured throughput for the scenario with TCP traffic, and proportional fair scheduler. Achieved performance ratio 0.4 0.6 0.8 1.0 1 2 3 4 5 6 7 8 9 10 11 12 User ID (b) Achieved performance ratio (predicted vs. measured MAC throughput) for the scenario with TCP traffic, and proportional fair scheduler. Achieved performance ratio 0.4 0.6 0.8 1.0 1 2 3 4 5 6 7 8 9 10 11 12 User ID (c) Achieved performance ratio (predicted vs. measured MAC throughput) for the scenario with UDP traffic, and round robin scheduler.
A simple linear regression model Machine learning techniques for performance estimation Scenario with 4 femtocell scenario and 3 users per femto and bandwidth of 2 MHz TCP traffic and proportional fair scheduler This method results in poorer prediction average both in terms of median error, as well as the variability of the results The increased amount of information available for the predictor results in both improved median prediction performance, as well as substantial reduction in the magnitude of the outliers
A simple linear regression model Figure Machine learning techniques for performance estimation Achieved performance ratio 0.4 0.6 0.8 1.0 1 2 3 4 5 6 7 8 9 10 11 12 User ID (c) Achieved performance ratio for the linear regression method with 10% of samples and the stratified sampler. Achieved performance ratio 0.6 0.7 0.8 0.9 1.0 1 2 3 4 5 6 7 8 9 10 11 12 User ID (c) Achieved performance ratio for the Kohonen regression method with 10% of samples and the random sampler. Achieved performance ratio 0.2 0.4 0.6 0.8 1.0 1 2 3 4 5 6 7 8 9 10 11 12 User ID (c) Achieved performance ratio for the PPR regression method with 70% of samples and the stratified sampler.
The initial results obtained here show that advanced regression techniques can result in good performance predictions In future work we plan to investigate an active sampling based on graph coloring