Deep Reinforcement Learning for Resource Allocation With Mixed Traffic in NOMA System

The non-orthogonal multiple access (NOMA) system with mixed traffic has emerged as a promising solution to satisfy the requirements of various services while mitigating the spectrum scarcity issue. In this paper, we investigate the resource allocation problem with mixed traffic in an uplink NOMA system, which consists of real-time (RT) users with strict quality-of-service (QoS) requirements on rate and latency, and best-effort (BE) users that aim to achieve a balance between throughput and fairness. In order to account for the diverse requirements of different users, we propose a utility function and formulate the resource allocation problem. To address this mixed integer non-convex problem, a deep reinforcement learning (DRL) algorithm is proposed which consists of two parts. Specifically, we design a deep Q network (DQN) to generate subcarrier assignment policies and a deep deterministic policy gradient (DDPG) network to generate power allocation policies. Simulation results demonstrate that the proposed algorithm outperforms existing algorithms in the proposed utility function.