Deep multi-agent reinforcement learning for resource allocation in NOMA-enabled MEC

Abstract

Non-orthogonal multiple access (NOMA) and mobile edge computing (MEC) are being considered as promising technologies to address the stringent demands of the emerging fifth generation (5G) networks. This paper investigates the resource allocation problem in NOMA-enabled MEC system for multiple users, by joint optimization of power and computation resources to enhance effective throughput of the system. Because of the severe non convexity of the problem, a decentralized multi-agent reinforcement learning (MARL) scheme is proposed, where each user to base station (U2B) link acts as an agent, and collectively interacts with the network environment, in order to maximize the objective function under limited power and computation resource constraints. Simulation results demonstrate that the proposed MARL scheme results in considerable improvement in the effective throughput of the system, which is comparable to the optimal results derived from the exhaustive optimal search, with significantly low overhead.