Diffusion-enhanced multi-agent reinforcement learning for adaptive interference control in bilateral cognitive radar games

In complex electromagnetic environments, adaptive signal-emitting systems operate under partial observability, asymmetric information, and dynamically evolving spectrum interactions, posing significant challenges to interference-aware decision-making. To address these challenges, this paper formulates the signal interaction process as a partially observable stochastic game and proposes a Bilateral Intelligent Game Policy Optimization Architecture (BIGPOA). Within this framework, a two-stage Diffusion-Enhanced Multi-Agent Proximal Policy Optimization (Diff-MAPPO) algorithm is developed. In the online stage, Multi-Agent Proximal Policy Optimization (MAPPO) with self-play training is employed to learn interactive behaviors between sensing and interference agents. In the offline stage, a diffusion model refines the learned policies through trajectory-level distribution regularization, while Double Deep Q-Network (DDQN)-guided optimization promotes high-value decision generation. Experimental results demonstrate that Diff-MAPPO achieves faster convergence, higher interference effectiveness, and improved energy efficiency compared with representative reinforcement learning baselines, while maintaining robust performance under dynamic and uncertain signal environments.