Low-Complexity Adaptive Digital Predistortion with Meta-Learning based Neural Networks

Alexis Falempin (1) , Rafik Zayani (2) , Jean-Baptiste Dore (1) , Emilio Calvanese Strinati (3)

1 DSYS - Département Systèmes
2 Université de Carthage - University of Carthage
3 CEA-LETI - Commissariat à l'énergie atomique et aux énergies alternatives - Laboratoire d'Electronique et de Technologie de l'Information

In this paper, we study a meta-learning based neural network (NN) model to enhance the energy-efficiency related to power amplification in wireless communication systems. Specifically, we introduce a low-complexity adaptive solution to perform digital predistortion (DPD) for power amplifiers (PAs) using neural networks. Thus, we design a dedicated NN architecture to derive predistortion functions using few neurons. Moreover, we develop a meta-learning training approach that allows better generalization and faster adaptation, over time-varying PAs, compared to classical DPD architectures. Thereby, we propose a new approach to realize an adaptive digital predisorter based on meta-learning using few samples for online calibration. A dedicated architecture allows to achieve low-complexity while meta-learning permits adapting to most parameters change in the system. Through the simulation results, we have shown that the developed meta-learning based NN DPD can offer a meta-trained DPD function, i.e., trained offline, that can provide satisfying performance for different PA models. Contrary to classical DPD architectures, the performance of our meta-trained DPD can be improved through only few gradient steps and over few samples during online calibration, achieving excellent performance with moderate complexity.

Domains: Signal and Image processing; Machine Learning [stat.ML]

Alexis Falempin, Rafik Zayani, Jean-Baptiste Dore, Emilio Calvanese Strinati. Low-Complexity Adaptive Digital Predistortion with Meta-Learning based Neural Networks. 2022 IEEE 19th Annual Consumer Communications & Networking Conference (CCNC), IEEE, Jan 2022, virtual event, United States. pp.453, ⟨10.1109/CCNC49033.2022.9700529⟩. ⟨cea-03837858⟩