Reinforcement Learning-Based IRS Phase Optimization in MIMO Systems

This MATLAB script implements a reinforcement learning (RL) approach to optimize IRS phase configurations in a MIMO wireless system. The implementation features a basic MIMO setup with a 16-element IRS operating at 12 GHz (mid-band frequency). Using the policy gradient method with a two-layer neural network, it learns optimal phase shifts while considering user mobility and Rician fading channels. The system models both direct and IRS-reflected paths, incorporating realistic path loss and channel conditions. The learning progress is visualized through a cumulative reward history plot, where rewards are computed based on achievable channel capacity.