Energy-Efficient Dynamic-Grouped-Adaptive Connected Hybrid Precoding with DRL in Cell-Free Massive MIMO Systems

In this paper, for mmWave cell-free massive MIMO systems, an energy efficiency optimization scheme that integrates a dynamic-grouped-adaptive (DGA) connected hybrid precoding at access points (APs) with deep reinforcement learning (DRL) is proposed. The proposed DGA-connected architecture divides antennas into groups and supports adaptive activation of radio frequency chains and phase shifters using switches. To further address the dynamic collaborative transmission challenges among APs with this architecture, we propose an action-embedded multi-agent soft actor-critic (AE-MASAC) algorithm incorporating a Transformer-based multi-head attention mechanism to facilitate collaborative behavior among AP agents. The simulation results show that the DGA-connected architecture can significantly reduce hardware power consumption compared to fully-connected and sub-connected structures. The AE-MASAC algorithm has advantages in convergence speed and scalability over traditional centralized DRL and distributed MADRL methods, providing a general framework for high-dimensional joint optimization problems in dynamic wireless communications systems.