HadouQen: Adaptive AI Agent Using Reinforcement Learning in Street Fighter II: Special Champion Edition

Abstract

This study presents the development of an AI agent trained using Proximal Policy Optimization (PPO) to compete in Street Fighter II: Special Champion Edition. The agent learned optimal combat strategies through reinforcement learning, processing visual input from frame-stacked grayscale observations (84 × 84 pixels) obtained through the OpenAI Gym Retro environment. Using a convolutional neural network architecture with carefully tuned hyperparameters, the model was trained across 16 parallel environments over 100 million timesteps. The agent was tested against M. Bison, the game's final boss and most challenging opponent, across 1,000 consecutive matches to evaluate performance. Results showed exceptional performance with a 96.7%-win rate and an average reward of 0.912. Training metrics revealed a healthy learning progression, showing steady improvement in average reward per episode, decreased episode length indicating more efficient victories, and stable policy convergence. The findings also demonstrate the effectiveness of PPO-based reinforcement learning in mastering complex fighting game environments and provide a foundation for future research in competitive game-playing agents capable of human-level performance in fast-paced interactive scenarios.