Machine Learning for IRS-Assisted MU-MIMO Communications with Estimated Channels

The reflective elements on an intelligent reconfigurable surface (IRS) can be tuned to improve the propagation environment and, in turn, the system performance, provided reliable channel information. However, IRS also brings challenges in terms of channel estimation by creating a very large dimensional channel between the IRS and the transmitter/receiver. A channel size being proportional to the number of users, user equipments (UE) antennas, IRS phase shifters, base station (BS) antennas, in an IRS-aided multi-user (MU) multiple-input-multiple-output (MIMO) system. Estimating such a large channel implies a huge amount of training overheads when all IRS phase shifters are passive, or requires extra power consumption and prohibitive hardware complexity, when IRS phase shifters are active. To tackle these challenges, this paper proposes a practical channel estimation process, based on a realistic codebook design constraint, for optimizing the IRS reflective elements in a MU-MIMO scenario. More specifically, we propose two novel machine-learning based algorithms (i.e. a deep supervised and a deep reinforcement) for optimizing the reflective elements of a typical passive IRS as well as a reliable channel estimation technique for IRS. Deep Supervised network uses exhaustive search to try every reflection pattern to train the network while the reinforcement network uses Q-learning to get the best reward. Our two algorithms can use the imperfect estimated channel knowledge to optimize the IRS in terms of sum-rate or minimum rate among all users. Simulation results show that our practical algorithms can achieve sum-rate and minimum rate performances close to the theoretical ones.