WSEE Optimization of Cell-Free mMIMO Uplink Using Deep Deterministic Policy Gradient

Cell-free (CF) massive multiple-input multiple-output (mMIMO) systems are being actively investigated. We optimize the weighted sum energy efficiency (WSEE) for the uplink of a CF mMIMO system by using deep deterministic policy gradient (DDPG) technique, which does not require labeled data. We formulate the WSEE optimization as a reinforcement learning problem. We first show the limitations of the reward function and states being used in the baseline DDPG technique from the existing CF state of the art. We construct improved reward functions and states, and then propose two novel DDPG variants which have better sampling and exploration capabilities than the baseline DDPG technique. The proposed variants provide 4% WSEE gain, and
$5\times $
faster convergence.