Intelligent Decision-Making for Asynchronous Dynamic Orthogonal Networking Based on DO-QMIX Algorithm

Abstract

In order to intelligently select the frequency points of each subnet in the asynchronous dynamic orthogonal networking (ADON), we propose the QMIX algorithm based on dataset aggregation and options architecture (DO-QMIX). Joint reward is maximized to mitigate the problem of partially observable environment. Specifically, we first pre-train the network by dataset aggregation (DAgger) to improve the sample utilization. Then, we fine-tune the policy via experiences generated by options architecture (OA) to avoid getting trapped in local optima. Numerical results show that the proposed DO-QMIX outperforms the comparison algorithms in the three complex electromagnetic environments.