RL-Based Digital Predistortion for Wideband Power Amplifiers

Power Amplifiers (PAs) are an integral part of any wireless communication chain. They increase the power of the transmitted signal, albeit non-linearly at certain power levels, to enable long-range transmission. This paper showcases a novel approach based on Reinforcement Learning (RL) that performs Digital Predistortion (DPD) for Radio Frequenccy (RF) PAs. The RL-agent predistorts the desired signal by manipulating the complex form of the signal to proactively compensate for the non-linearity effect on the PA output. The proposed model accounts for the memory effect of the PA. The PA in this work was modeled after a GMP whose coefficients were derived from an actual PA. Results show that the RL-DPD achieves a significant performance boost, reducing Error Vector Magnitude (EVM) by nearly 80% compared to Generalized Memory Polynomial (GMP) DPD, and improving Normalized Mean Square Error (NMSE) by over 13 dB. Additionally, RL-DPD outperformed the existing machine learning based ARVTDNN and RVTDCNN in NMSE by approximately 4 dB, while also delivering a 0.6 dB and 0.34 dB improvement in Adjacent Channel Power Ratio (ACPR) respectively. With these results, the effectiveness of the RL-DPD technique was thoroughly evaluated, positioning this work as a base for further investigation.