street-fighter-ai

Why we included this project

If you are learning deep reinforcement learning by doing, this project is a rare find: it walks you from raw RGB pixels on the screen all the way to a trained policy that beats the final boss of Street Fighter II: Special Champion Edition. Built on Stable-Baselines3 and OpenAI Gym Retro, the agent makes every decision from the raw pixel values alone, with no hand-crafted feature extraction to lean on. The repository includes model weights from several training stages and TensorBoard logs, so you can watch the strategy sharpen and rerun the whole pipeline on Python 3.8. The authors also openly flag that the win rate of 100 percent in the provided save state comes with overfitting to that single scenario, which makes for an honest lesson about evaluating RL agents beyond one fixed test. The codebase is small enough to read in full, which makes it a better companion for coursework or solo study than a drop-in tool.