Reconfigurable Intelligent Surface Enabled Over-the-Air Federated Learning: A Deep Reinforcement Learning Approach

Federated learning (FL) is a cost-efficient and privacy-preserving distributed learning framework. To enhance communication efficiency during FL model aggregation, over-the-air computation has been introduced, which exploits the superposition property of wireless channels to facilitate concurrent local weight communication and global model computation. However, over-the-air computation may introduce signal distortion and global aggregation errors, and may face hindrances such as blockages between devices and the parameter server. This paper explores reconfigurable intelligent surface (RIS)-enabled over-the-air FL to address these issues. We present a novel deep reinforcement learning (DRL) algorithm for designing RIS phase shifts to minimize mean squared error (MSE) during over-the-air computation. Simulation results demonstrate lower MSE and higher test accuracy yielded by the proposed DRL algorithm compared to existing model-based approaches.