Wireless Channel Prediction for Multi-user Physical Layer with Deep Reinforcement Learning

In this paper, we consider a reinforcement learning (RL) based multi-user downlink communication system. An actor-critic based deep channel prediction (CP) algorithm is proposed at the base station (BS) where the actor network directly outputs the predicted CSI without channel reciprocity. Different from the existing methods which either require the perfect channel state information (CSI), or estimate outdated CSI and set strict constraints on pilot sequences, the proposed algorithm has no such premised knowledge requirements or constraints. Deep-Q learning and policy gradient methods are adopted to update the parameters of the proposed prediction network, with the objective of maximizing the overall transmission sum rate. Numerical simulation results and the complexity analysis verify that the proposed CP algorithm could beat the existing traditional and learning based methods in terms of sum rate over different channel models and different numbers of users and antennas.