AlphaZero-Style Game Agent

Year

2024

Tech & Technique

PyTorch, CUDA, Monte Carlo Tree Search, Policy-value network

Description

AlphaZero-style game agent: self-play reinforcement learning with Monte Carlo Tree Search (MCTS) guided by a policy-value network, in PyTorch with CUDA. Extended Robert Förster's AlphaZero tutorial. Reached an 81% win rate vs rule-based opponents.

Key Facts:

🌲 Monte Carlo Tree Search guided by a policy-value network.
⚙️ PyTorch and CUDA.
🏁 81% win rate vs rule-based opponents.

Credit: the base implementation comes from Robert Förster's AlphaZero tutorial.

My Role

Personal project — extended Robert Förster's AlphaZero tutorial

🌲 Worked with MCTS and a policy-value network in PyTorch (CUDA).
🏁 Reached an 81% win rate vs rule-based opponents.

Go Board

Illustrative photo (third-party image).

Separate Project: ChessAI

Screenshot of my separate Pygame ChessAI project — not the AlphaZero-style agent.

Generic CNN Diagram

Illustrative convolutional-network diagram (third-party image) — not this project's policy-value network.

Othello Agent Figure (third-party)

Third-party figure of a learned Othello agent — not from this project.

AlphaZero Chess Concepts (third-party)

Third-party figure on chess concepts learned by AlphaZero — not from this project.