Examinator: Thomas Eriksson, Inst för elektroteknik
The purpose of this thesis is to explore the possibility of using machine learning (ML) algorithms for the behavioral modeling of power amplifier (PA). This thesis is an extension of the previous masters thesis work  which compares the performance of ML methods with memory polynomial (MP), generalized memory polynomial (GMP) and look-up table (LUT) methods for the PA modelling on single carrier and single band scenario. We expand it to multicarrier and multiband scenarios. The performance of MP and GMP as the baseline algorithms are compared with the performance of ML algorithms as neural network (NN), gradient boosting (GB), decision trees (DT) and linear regression (LR) in terms of normalized mean square error (NMSE) and adjacent channel error power ratio (ACEPR). The experiments are done with three different test scenarios for single carrier as a reference case, multi carrier in full band, and multi carrier in separated band/carrier. Experimental results show that NN achieves the best performance in all of the test scenarios except for the separated multi-carrier scenario and GB also gives the best performance in one of the separate multi-carrier scenario but in most scenarios the performance is worse than MP and GMP results. DT gives poor performance at all and finally LR fails to correctly predict the test output signal as expected that the PA modeling can be considered the non-linearity problem. Finally, computational complexity analysis of ML algorithms is given with the corresponding performance results of ML algorithms in this study.