A review of reinforcement learning approaches for pursuit-evasion games

As a special type of dynamic game, Pursuit-Evasion Games (PEGs) have expanded their application range from initial military confrontations to areas such as navigation control and aerospace, demonstrating broad applicability and significant value in addressing a wide array of modern complex decision-making problems. Traditional optimal control methods based on differential game theory are classic approaches to solve PEG problems. However, these methods often struggle to perform well in complex environments, nonlinear systems, and situations involving highly uncertain participant behaviors. In recent years, rapidly developing Reinforcement Learning (RL) techniques has provided new avenues for PEG research. RL is capable of adapting to environmental changes through efficient online computation and feedback-driven learning, exhibiting strong generalization capabilities. Therefore, this survey presents a detailed and systematic review of PEG research based on RL methods. First, it classifies and discusses key RL algorithms and theoretical foundations in PEGs according to different forms of strategy learning. Then, it summarizes typical application scenarios, including tactical combat, unmanned systems control, and spacecraft interception, demonstrating the potential and effectiveness of RL in addressing real-world challenges. Finally, the survey explores current challenges and future opportunities in applying RL to PEGs, with the aim of promoting further research on more effective and practical solutions.