AI Based Low Complexity Design for Digital Pre-Distorter for Next Generation Wireless Systems

Power amplifiers (PA) are the essential part of the wireless communication systems but generally exhibit non-linearity at the high voltage input. Non-linearity in the orthogonal frequency division multiplexing (OFDM) based transceiver can cause severe distortions for the both in-band and out-band signal because of the high peak to average power ratio (PAPR). As the system bandwidth goes higher, PA exhibit memory effect. Non-linearity and memory effect is usually handled efficiently with digital pre-distorter (DPD). DPD is modeled with Volterra series kind of polynomials specially generalized memory poly-nomial (GMP). Estimation of the coefficients of the GMP is non-trivial and involves lot of complexity when the order and the memory of the GMP are significantly high. In this paper, we propose AI-methods to select the top-P features of the GMP based DPD out of total $N$ features (P < < N). We propose a modified concrete selector based AI model, which selects top-$P$ features during the training of the AI model. We also propose a perturbation based method, which perturbs the input and based on its effect on the output obtained from the trained model, finds the most suitable top- $P$ features. At last, we show the performance of our methods by calculating the performance metrics such as error vector magnitude (EVM) and adjacent channel leakage ratio (ACLR). We show that for the GMP with N = 602 features, AI based selected top-50 GMP features performs similar to the GMP with all 602 features, and thus the complexity is reduced by more than 99%.