基于深度强化学习的协作非正交多址接入网络鲁棒安全传输

For a cooperative non-orthogonal multiple access (NOMA) network over power line channels in the presence of untrusted relays, this paper first investigated a robust secure transmission problem by performing the optimal relay selection and optimizing the power between the confidential NOMA signal and the jamming signal. Considering the bounded channel uncertainties, a hierarchical deep reinforcement learning (DRL) scheme based on quantized channel state information (CSI) was proposed in order to maximize the system secrecy sum rate while guaranteeing the destination nodes’ quality of service requirements and the maxim transmit power constraint. In this proposed scheme, the joint optimization problem was decomposed into relay selection subproblem and power allocation subproblem, and then the deep Q-learning (DQL) method was adopted to learn the optimal action policy for each subproblem in the decomposed action spaces. In addition, the DRL states and actions were carefully designed based on the quantization interval index of CSI. Simulation results show that the proposed scheme has a great ability in dealing with the curse of dimensionality with less computation. In addition, it can adaptively adjust the action policies to cope with the changes of network scale, which means that the proposed scheme has good scalability and generalization performance.