When mastering complex games, human players form abstractions of concepts and strategies that obey the game rules, which helps improve performance. These abstractions serve both to explain opponents' moves and to guide one’s own decisions. Understanding an opponent’s strategy is essential for improvement, yet it usually demands considerable time and effort.
This paper introduces a human‑like cognitive approach for training reinforcement‑learning agents to synthesize learned strategies and policies into executable procedures based on sequences of gameplay actions. We present an automatic program‑learning pipeline that can infer concise strategy programs from chess game logs and trajectories in a grid‑world environment.
Experiments on chess and a grid‑based task demonstrate that the synthesized programs generate effective actions during play and can be learned solely from existing gameplay data.
Review