State-Based Potential Game Approach for Anti-Jamming Channel Access via Reinforcement Learning

This paper investigates the distributed anti-jamming channel access in multi-user communication networks. In anti-jamming wireless communication networks, the dynamic characteristics of external jamming, the absence of centralized network control, and the localized impact of interference on network nodes create significant hurdles for anti-jamming strategies. To address these challenges, we model the anti-jamming problem using a state-based potential game framework, theoretically guaranteeing convergence. We then propose a Neighborhood-Aware Cooperative Multi-Agent Reinforcement Learning (NAC-MARL) algorithm. This algorithm effectively mitigates external malicious jamming while minimizing mutual interference among users by facilitating local information exchange, enhancing decision-making, and enabling adaptive responses to dynamic jamming conditions. Experimental results demonstrate that the proposed NAC-MARL algorithm achieves a 13.1% increase in network throughput compared to independent Q-learning and a 43.1% improvement over the SAP method under dynamic sweeping jamming conditions. Additionally, the algorithm reduces the average congestion level across various user scenarios, ensuring stable communication in dynamic spectrum environments.