AI Learns Action Duration in Fighting Games

A new reinforcement learning framework allows agents to learn both the action and its duration in fighting games like Street Fighter II. Traditional RL agents act at fixed intervals (every frame or N frames), granting frame-perfect reflexes or reducing computational cost but lacking adaptability. The proposed method jointly predicts action and duration, enabling dynamic responsiveness. Implemented in the open-source FightLadder environment, agents trained against scripts can adjust reaction timing per situation. The research, published on arXiv (2605.20911), addresses the challenge of real-time decision-making in fast-paced games.