Monte Carlo Tree Search and Neural Networks for Hex

This project implements an On-Policy Monte Carlo Tree Search (MCTS) system combined with reinforcement learning and a neural network to play the board game Hex.

Project Overview

• Goal: Train an intelligent agent to play Hex using MCTS, reinforcement learning, and a policy network.
• MCTS: Guides the agent's decisions by simulating games and using the results to train the policy network.
• Reinforcement Learning: Updates the policy network after each game based on the results of many simulated games.
• Neural Network (ANET): Learns to predict optimal moves from game states and is periodically updated throughout training.

Key Features

• Hex Game: Simulate and play the board game Hex using a k x k grid (3 ≤ k ≤ 10).
• MCTS: Performs simulations to evaluate and backpropagate move values, refining the policy over time.
• Training and Competitions: The agent improves by playing games, and periodically saved models compete in the Tournament of Progressive Policies (TOPP).
• On-Policy Training: The same policy network is used for decision-making during simulations and training.