A Multi-Agent DRL-Based Power Allocation Mechanism for Multi-Cell NOMA Networks

Abstract

To effectively address the power allocation (PA) problem to maximize the energy efficiency (EE) for a non-orthogonal multiple access (NOMA) system, a novel action mapper alongside a multi-agent deep reinforcement learning (MADRL)-based algorithm is designed in this paper. The action mapper discretizes and merges similar actions into one action so that the action space can be significantly reduced. Then, it is integrated with a multi-agent proximal policy optimization (MAPPO) algorithm to efficiently perform the PA task. Our proposed MADRL-based algorithm with the reduced action mapper is able to find the sub-optimal solution according to the current environmental condition. Supported by our simulation results, our proposed mechanism can significantly improve the EE as compared to the closely related approaches.