Fast and Scalable Intelligent Reflecting Surface Beamforming Using Hybrid Learning Approaches

This paper presents a novel hybrid optimization framework for intelligent reflecting surface (IRS)-assisted beamforming, which synergistically integrates reinforcement learning (RL) and cross-entropy (CE) methods. The proposed approach addresses the key limitations of conventional techniques, such as high computational complexity and slow convergence in large-scale IRS deployments. By combining the global exploration capability of RL with the precise local refinement of CE, the framework efficiently optimizes IRS phase shifts to maximize the signal-to-interference-plus-noise ratio (SINR). The bidirectional interaction between the RL and CE components enables adaptive beamforming in dynamic environments, while the joint optimization of transmit and reflective beamforming eliminates the need for alternating optimization. Simulation results demonstrate that the proposed method outperforms state-of-the-art techniques, achieving a SINR of 35 dB with 512 IRS elements, a sum rate of 42 bps/Hz, and 40% faster convergence compared to deep RL approaches. Furthermore, the framework exhibits robust performance with less than 2.5 dB SINR degradation under 30% channel state information errors and supports user mobility up to 10 m/s. These advancements, combined with an energy efficiency of 25 bits/Joule/Hz, establish the practical viability of the proposed solution for future 6G wireless networks.