Eugene Peng

Back

Connect4 AlphaZero

View on GitHub

Overview#

A reinforcement learning agent that masters the game of Connect 4 using the AlphaZero algorithm.

Combines Monte Carlo Tree Search (MCTS) with a dual-headed residual neural network. The large model (5 ResBlocks, 128 filters) beats Minimax depth-11 at 100% and placed top 10 on the Kaggle Connect X leaderboard using only ~50 MCTS simulations per move within Kaggle’s 2-second time limit.

Features#

  • Neural network: 5 Residual Blocks, 128 Filters, ~1.6 million parameters.
  • Training: 40,000 self-played games