Reinforcement Learning in Games

Abstract

This project applies reinforcement learning to train agents to play the classic fighting game Street Fighter 2. The project uses the OpenAI Gym Retro library to create an environment for running the game and providing a platform for training reinforcement learning models. Two state-of-the-art reinforcement learning algorithms are employed for training agents in this environment: Proximal Policy Optimization (PPO) and Recurrent Proximal Policy Optimization (RecurrentPPO). To optimize training results for the specifics of Street Fighter 2, a novel reward function is designed. The reward function rewards agents for defeating their opponent, using special moves, balancing offensive and defensive actions, and penalizes them for being knocked out or failing to execute actions. Through extensive training of PPO and RecurrentPPO agents with the custom reward function, the project aims to develop agents that can play Street Fighter 2 at a superhuman level. The trained agents will be evaluated based on their ability to defeat human players. The project seeks to demonstrate the potential of reinforcement learning for mastery of complex real-world games with partial observability. By proposing an optimal way to frame rewards for Street Fighter 2, the project also aims to provide insights that can benefit future work applying reinforcement learning to other complex games. In summary, this project applies reinforcement learning and develops new techniques to achieve human-level and super-human performance in the classic fighting game Street Fighter 2. The developed methodology and results will provide a foundation for expanded research on mastering complex real-world games through reinforcement learning.