简体中文 | English | Español

This project is an AI agent trained using deep reinforcement learning to beat the final boss in the game "Street Fighter II: Special Champion Edition". The AI agent makes decisions based solely on the game screen's RGB pixel values. In the provided save state, the agent achieves a 100% win rate in the first round of the final level (overfitting occurs, see the Running Tests section for discussion).

├───data
├───main
│   ├───logs
│   ├───trained_models
│   └───scripts
├───utils
│   └───scripts

The game configuration files are stored in the data/ folder, and the main project code is in the main/ folder. Within main/, the logs/ folder contains terminal/console outputs and data curves recording the training process (viewable with Tensorboard), while the trained_models/ folder contains model weights from different stages. These weights can be used for running tests in test.py to observe the performance of the AI agent's learned strategies at different training stages.

This project is based on the Python programming language and primarily utilizes standard libraries like OpenAI Gym Retro and Stable-Baselines3. The Python version used is 3.8.10, and it is recommended to use Anaconda to configure the Python environment. The following setup process has been tested on Windows 11. Below are console/terminal/shell commands.

# Create a conda environment named StreetFighterAI with Python version 3.8.10
conda create -n StreetFighterAI python=3.8.10
conda activate StreetFighterAI

# Install Python libraries
cd [parent_directory_of_project]/street-fighter-ai/main
pip install -r requirements.txt

# Run script to locate gym-retro game folder
cd ..
python .\utils\print_game_lib_folder.py

After the console outputs the folder path, copy it to the file explorer and navigate to the corresponding path. This folder contains the game data files for "Street Fighter II: Special Champion Edition" within gym-retro, including the game ROM file and data configuration files. Copy the Champion.Level12.RyuVsBison.state, data.json, metadata.json, and scenario.json files from the data/ folder of this project into the game data folder, replacing the original files (administrator privileges may be required). The .state file is a save state for the game's highest difficulty level, while the three .json files are gym-retro configuration files storing game information memory addresses (this project only uses [agent_hp] and [enemy_hp] for reading character health values in real-time).

To run the program, you will also need the game ROM file for "Street Fighter II: Special Champion Edition", which is not provided by gym-retro and must be obtained legally through other means. You can refer to this link.

Once you have legally obtained the game ROM file, copy it to the aforementioned gym-retro game data folder and rename it to rom.md. At this point, the environment setup is complete.

Note 1: If you want to manually capture save states and find memory variables in the game, you can use the gym-retro integration ui. Copy data/Gym Retro Integration.exe to the parent menu (two levels up, retro/ folder) of the aforementioned gym-retro game data folder.

Note 2: If you want to record videos of the AI agent's gameplay, you will need to install ffmpeg.

conda install ffmpeg

Once the environment is set up, you can run test.py in the main/ folder to test and experience the AI agent's performance at different stages of training.

cd [parent_directory_of_project]/street-fighter-ai/main
python test.py

Model weight files are stored in the main/trained_models/ folder. The default model used in test.py is ppo_ryu_2500000_steps_updated.zip, which has good generalization and is capable of beating the final level of Street Fighter II: Special Champion Edition. If you want to see the performance of other models, you can change the model_path variable in test.py to the path of another model file. The observed performance of the models at various training stages is as follows:

• ppo_ryu_2000000_steps_updated: Just beginning to overfit state, generalizable but not quite capable.
• ppo_ryu_2500000_steps_updated: Approaching the final overfitted state, cannot dominate first round but partially generalizable. High chance of beating the final stage.
• ppo_ryu_3000000_steps_updated: Near the final overfitted state, almost dominate first round but barely generalizable.
• ppo_ryu_7000000_steps_updated: Overfitted, dominates first round but not generalizable.

If you want to train your own model, you can run train.py in the main/ folder.

cd [parent_directory_of_project]/street-fighter-ai/main
python train.py

The project includes Tensorboard graphs of the training process. You can use Tensorboard to view detailed data. It is recommended to use the integrated Tensorboard plugin in VSCode to view the data directly. The traditional viewing method is listed below:

cd [parent_directory_of_project]/street-fighter-ai/main
tensorboard --logdir=logs/

Open the default Tensorboard service address http://localhost:6006/ in your browser to view interactive graphs of the training process.

This project uses open-source libraries such as OpenAI Gym Retro, Stable-Baselines3. The contributions of all the developers to the open-source community are appreciated!

Two papers that had a significant impact on this project:

[1] DIAMBRA Arena A New Reinforcement Learning Platform for Research and Experimentation The valuable summary of the experience in setting hyperparameters for deep reinforcement learning models in fighting games in this paper was of great help to the training process of this project.

[2] Mitigating Cowardice for Reinforcement Learning The "penalty decay" mechanism proposed in this paper effectively solved the "cowardice" problem (always avoiding opponents and not daring to even try attacking moves).