Digital Predistortion Model Extraction Using Reinforcement Learning Without Time-Domain Output Signal Acquisition

This article proposes a novel digital predistortion (DPD) model extraction technique for RF power amplifiers (PAs) using reinforcement learning (RL). Unlike conventional methods that extract model coefficients by comparing the input and output signals of the PA in the time domain, the proposed approach uses an RL agent that interacts directly with the PA and iteratively extract the DPD coefficients based solely on scalar performance metrics, such as adjacent channel power ratio (ACPR) or out-of-band (OOB) power, thereby eliminating the need for acquiring high-resolution time-domain output signals. To address the unique challenges of the DPD model extraction problem using RL, the deep deterministic policy gradient (DDPG) algorithm is specifically adapted by introducing a lightweight actor network and a single-step training strategy, substantially reducing computational complexity. Experimental results using 5G new radio (NR) signals with 256-QAM modulation at the 100- and 200-MHz bandwidths on a Doherty PA demonstrate that the proposed approach achieves effective linearization performance while offering strong application potentials in wideband scenarios.