Reinforce4J: A Reinforcement Learning Framework for Board Games

About The Project

Reinforce4J is a Java-based framework for building and training reinforcement learning agents to play classic two-player board games. This project provides a complete end-to-end implementation of a self-play training pipeline inspired by DeepMind's AlphaZero. It uses a Monte Carlo Tree Search (MCTS) algorithm to explore game states and generate training data for a neural network, which then acts as a powerful evaluator to guide the search. The framework is designed to be modular and extensible, allowing developers to easily add new games and experiment with different reinforcement learning models.

---

Key Features

• Complete Reinforcement Learning Pipeline: Implements a full self-play and training loop.
• Monte Carlo Tree Search (MCTS): Sophisticated game tree exploration for intelligent move selection.
• Neural Network Integration: Supports both TensorFlow and ONNX for model evaluation, allowing for flexibility in model deployment.
• Extensible Game Framework: Easily add new games by implementing a simple GameState interface. Comes with implementations for Tic-Tac-Toe and Connect Four.
• Modular and Configurable: Uses Guice for dependency injection, making the components highly decoupled and easy to configure.
• Comprehensive Testing: Includes a robust suite of unit tests to ensure code quality and correctness.

Technology Stack

• Core Framework: Java
• Deep Learning: TensorFlow, ONNX Runtime
• Dependency Injection: Google Guice
• Build Tool: Maven
• Testing: JUnit 5, Google Truth

---

Project Structure

The codebase is organized into the following main packages:

• core: Contains core abstractions like GameState, Player, and Outcomes.
• games: Implementations of specific games like TicTacToe and Connect4.
• evaluation: Provides different model evaluators, including TensorflowEvaluator and OnnxEvaluator.
• montecarlo: The core MCTS implementation, including TreeNode, TreeSearch, and expansion strategies.
• playing: Includes game simulation logic (PlayOutSimulator) and strategies for move selection.
• learning: Contains the reinforcement learning pipelines and model training execution logic.
• resources: Holds the Python scripts for training the neural network models.

---

Contributing

Contributions are welcome! If you'd like to contribute, please fork the repository and create a pull request. You can also open an issue with the tag "enhancement".

Possible areas for contribution include:

• Adding new games.
• Implementing more sophisticated MCTS strategies.
• Improving the neural network models.
• Adding support for other deep learning frameworks.