Abstract

A multi-agent deep reinforcement learning framework is proposed to address link level throughput maximization by power allocation and modulation and coding scheme (MCS) selection. Given the complex problem space, reward shaping is utilized instead of classical training procedures. The time-frame utilities are decomposed into subframe rewards, and a stepwise training procedure is proposed, starting from a simplified power allocation setup without MCS selection, incorporating MCS selection gradually, as the agents learn optimal power allocation. The proposed method outperforms both weighted minimum mean squared error (WMMSE) and Fractional Programming (FP) with idealized MCS selections.