Learning-Aided Resource Allocation for Pattern Division Multiple Access-Based SWIPT Systems

Abstract

In this letter, a learning-aided resource allocation scheme based on the constrained Markov decision process (CMDP) is proposed to improve the average network energy efficiency (EE) with the constrained quality of service (QoS) in the pattern division multiple access (PDMA)-based simultaneous wireless information and power transfer (SWIPT) system. In order to solve the formulated CMDP resource allocation problem, the Lagrange duality is adopted to transform CMDP into an unconstrained Markov decision process (MDP). Due to the instability of the practical system, the Deep Q Network (DQN)-based CMDP scheme is proposed to obtain the optimal solution. The simulation results verify the proposed scheme converges faster than the benchmark in terms of increasing average network EE.