Low-Complexity Adaptive DPD: From Online Optimization to Meta-Learning

In this paper, we investigate low-complexity and adaptive digital predistorsion (DPD) techniques in order to enhance wireless communication system latency, complexity and power consumption. Specifically, we introduce efficient meta-learning based solutions for time-varying power amplifier (PA) linearization that allow to reduce hardware complexity. Our first proposed solution consists in a meta-learning based neural network (NN) model that is capable to perform, offline, an optimal initialization of the NN based DPD. This leads to an efficient and fast adaptation of the DPD to the time-varying PA characteristics, i.e., few shots are needed during online calibration. Interestingly, we introduced a different approach, referred to as DPD NN Weights Selector (DPDNNWS), that offers the ability to approximate more accurately the current PA characteristic. These solutions have been compared in terms of complexity and performance w.r.t Error Vector Magnitude (EVM) and Adjacent Channel Leakage Ratio (ACLR). Performance of the proposed approaches are benchmarked with the conventional learning based DPD approach. To offer low-complexity DPD, a dedicated NN structure is designed to derive DPD functions using few neurons. Our numerical simulations demonstrated that our proposed DPDNNWS and meta-learning approaches provide satisfying results when used online for several PA models. Contrary to conventional DPD, these approaches can be used to reduce hardware complexity implementation and data usage during online adaptation. Indeed, through numerical evaluation, it appears that meta-learning can reach satisfying performance using only 3000 IQ symbols during online adaptation. Besides, the DPDNNWS approach exhibits performance close to a linear power amplifier. Both solutions allow to achieve excellent performance compared to the state of the art solutions.