A Policy Style Recognition-Based Algorithm for Robust Policy Generation in Multi-Agent Adversarial Games

Current multi-agent reinforcement learning methods face challenges such as environmental nonstationarity and poor policy adaptability, often leading to locally optimal solutions tailored to the training environment. To address these issues, this paper proposes a policy generation algorithm named policy style recognition multi-agent deep deterministic policy gradient. The algorithm quantitatively describes and categorizes policy styles, using a supervised learning-trained network to recognize opponents’ policy styles in real time. This allows agents to dynamically adjust strategies during adversarial tasks, significantly enhancing robustness. Additionally, the strategy generation phase incorporates subgame perfect Nash equilibrium solutions from finite repeated games for further optimization. Experimental results show that the proposed method outperforms other existing algorithms, maintaining stable and superior performance across five different opponent strategies, even when facing previously unseen policies. These experiments demonstrate that the proposed approach greatly improves the stability and robustness of adversarial strategies.