Trained with Reinforcement Learning / PPO / PyTorch / Stable Baselines, inspired by @nicknochnack.
- Use health-based reward function instead of score-based so the agent can learn how to defend itself while attacking the enemy. It makes a defensive strategy to win the game. You may want to modify the function to penalize the time spent for a more offensive strategy.
- Reduce the action space from MultiBinary(12) (4096 choices) to Discrete(14) to make the training more efficient. You may want to add combos to make the fight more interesting.
- Make the episode one fight only instead of best-of-three. It helps value estimation.
- Preprocess the environment with AtariWrapper (NoopReset, FrameSkip, Resize, Grayscale) to reduce input size.
- Use multiprocessing for faster training and bypass OpenAI Gym Retro limitation on one environment per process.
- 100% win-rate against the enemy
- Much HP remaining after the fight
- Trained model is saved in latest_model.zip
- Training and Testing scripts can be viewed in StreetFighter.ipynb
