Action-masked reinforcement learning with decision-tree distillation for ship air defense

Objectives To address the problems of numerous incoming targets, short engagement windows, limited weapon resources, and the poor interpretability of deep reinforcement learning policies in shipborne air self-defense missions, this study investigates an autonomous interception decision-making method that balances action executability and policy credibility.Methods Based on a red-blue confrontation simulation environment, the shipborne air self-defense process is modeled as a Markov decision process, and a Proximal Policy Optimization (PPO) algorithm based on deep reinforcement learning is adopted to learn air-defense engagement policies; furthermore, procedural knowledge and decision-making knowledge are transformed into state-dependent action masks to suppress invalid and illegal actions, and the trained neural-network policy is distilled by means of a decision tree. Results Prior knowledge is transformed into a constrained action space to enable efficient PPO-based policy learning, and policy interpretability is achieved through decision-tree rule extraction. ConclusionsThe proposed method provides an implementation pathway for executable and auditable autonomous decision-making in highly dynamic and strongly constrained anti-missile scenarios.