Energy-efficient resource allocation based on deep reinforcement learning for massive MIMO-NOMA-SWIPT system

Abstract

Massive multiple-input multiple-output (MIMO) and non-orthogonal multiple access (NOMA) can improve the energy efficiency (EE) of simultaneous wireless information and power transfer (SWIPT) systems by jointly optimizing the power allocation coefficient and the time slot handover coefficient. However, improper allocation of resources severely restricts the improvement of EE and substantially elevates communication costs of SWIPT. To address these issues, this study proposes a joint resource allocation algorithm for optimizing user scheduling, power allocation, and power splitting based on distributed multi-agent double deep Q-network and double multi-agent deep deterministic policy gradient (MADDQN-DMADDPG) network. Specifically, a double deep Q-network (DDQN) is used to optimize user scheduling by decoupling the action selection from action evaluation, which resolves the Q-value overestimation and mitigates multi-user interference. Then, to improve the training stability, a deep deterministic policy gradient (DDPG) is utilized to optimize the power allocation and power splitting by leveraging its ability to handle continuous action spaces. Moreover, we introduce the distributed multi-agent learning to bolster the learning capabilities of DDQN and DDPG, ensuring accuracy and efficiency of resource allocation. Simulations demonstrate that the proposed algorithm can significantly improve the overall EE of system with fast convergence speed.