With the rapid development of modern power systems, traditional scheduling and reinforcement learning methods often fail to meet stringent Quality of Service (QoS) demands for low latency, high reliability, and stable bandwidth under large-scale bursty traffic. To address this problem, we propose a QoS-driven routing optimization approach based on Adversarial Reinforcement Learning, referred to as Adversarial Critic-Cooperative Actor (ACCA). By introducing adversarial agents that model worst-case perturbations, ACCA establishes a multi-agent game framework that enhances policy robustness and adaptability in dynamic network environments. Furthermore, a multi-dimensional state representation and a QoS-aware cost function are designed to capture metrics such as delay, bandwidth utilization, queue length, and packet loss. Experiments demonstrate that ACCA outperforms traditional routing protocols and standard reinforcement learning algorithms in terms of end-to-end delay, load balancing, and throughput, thereby providing an effective solution for QoS assurance in intelligent power communication networks.
He et al. (Thu,) studied this question.