Deep Reinforcement Learning-Based Strategies for Counteracting Parallel Approaching Guidance Laws in Orbital Pursuit-Evasion Games

Artificial Intelligence (AI) methods have been preliminarily applied to the study of spacecraft orbital pursuit-evasion problems. However, the lack of comparative analysis with traditional classical strategies has limited the full demonstration of the advantages of AI algorithms and has not allowed for a deep understanding of their potential drawbacks. In this paper, we construct a pursuit-evasion game model that incorporates practical constraints, such as maneuverability limitations and regional constraints. Drawing on the concept of parallel approach from classical missile guidance methods, we design a quasi-parallel approach control law for the pursuer based on the relative position and velocity direction metrics of the pursuer and evader. Using the Deep Deterministic Policy Gradient (DDPG) algorithm as the fundamental framework, we develop a training environment for intelligent evasion strategies. We conduct two-dimensional and three-dimensional simulation experiments under a regional guidance scenario to explore the advantages and countermeasures of AI strategies compared with quasi-parallel approach strategies. Additionally, we verify the guiding role of the reward conditions through a targeting experiment in a two-dimensional plane.