Street Fighter Reinforcement Learning

This repository contains code for training a reinforcement learning model to play Street Fighter II: Special Champion Edition using the Proximal Policy Optimization (PPO) algorithm.

Setup

Remeber to set up the game from the roms folder using the terminal To run the code, follow the instructions below:

pip install gym gym-retro
pip install opencv-python
pip install stable-baselines3[extra] optuna

Environment Setup

• The StreetFighter class is implemented as a custom environment using the Gym interface.
• Grayscaling, frame delta, and frame resizing are applied to preprocess observations.

Filtered Action Setup

• The StreetFighter environment uses retro.Actions.FILTERED to filter the action space.
• This helps focus on relevant actions for optimal training.
• Random Sampling of Actions
• https://github.com/FruitPnchSamuraiG/StreetFighter/assets/146465657/b182aa1c-035b-4718-b02d-e55c5346aeb5

Reward Function

• The reward function is set to the in-game score.
• The step method reshapes the reward function based on the score delta.

Training

• Hyperparameters are optimized using Optuna, and the best model is saved.
• The TrainAndLoggingCallback class defines a callback to save the model during training.
• The Train Model section demonstrates how to train the model using the best hyperparameters.

Evaluation

• The trained model is evaluated using the evaluate_policy function. https://github.com/FruitPnchSamuraiG/StreetFighter/assets/146465657/bbcf38d0-020c-4cfa-98ea-7d4e42186f6b

Test the Model

• The Test out the Model section demonstrates how to use the trained model to play the game. https://github.com/FruitPnchSamuraiG/StreetFighter/assets/146465657/3c435c54-ce85-445f-9d76-931c7cefc926

Files and Directories

• logs/: TensorBoard logs during training.
• opt/: Saved optimized models.
• train/: Checkpoints saved during training.

Usage

Follow the Jupyter notebook StreetFighter_Tutorial.ipynb to understand the code execution flow.

Happy gaming and training!