A Lightweight RL-Driven Deep Unfolding Network for Robust WMMSE Precoding

Weighted Minimum Mean Square Error (WMMSE) precoding can achieve near-optimal weighted sum rate (WSR) in MU-MIMO-OFDM systems, but it suffers from high computational complexity and severe dependence on accurate channel state information (CSI). This letter proposes a lightweight reinforcement learning (RL)-driven deep unfolding (DU) network (RLDDU-Net) to overcome these limitations. Specifically, its DU module maps a wideband stochastic WMMSE algorithm, which maximizes the ergodic WSR under imperfect CSI, into a deep-learning framework. Approximation techniques exploiting beam-domain sparsity and subcarrier correlation are also adopted to significantly reduce complexity and accelerate convergence. The RL module dynamically generates compensation matrices to mitigate approximation errors and adjusts DU network depth online, thereby enhancing both flexibility and performance. Simulation results demonstrate that RLDDU-Net outperforms existing baselines in terms of ergodic WSR, computational efficiency, and convergence speed under imperfect CSI.