ML-Agent Boxing

January 01, 2020

Boxing Agents gif

I threw this project together over the course of a week as an entry in Unity’s first challenge for their new ‘Ml-Agents’ machine learning workflow. I had tossed around the idea of using machine learning to train AI agents in fighting games, but I had never gotten around to it until then. The agents learn via Reinforcement Learning and are given a simple policy that positively rewards landed punches on the opponent, and negatively awards taking damage. Other behaviors such as successful blocking, and advancing on the opponent are rewarded as well.