Reproduce “Train on Small, Play the Large: Scaling Up Board Games with AlphaZero and GNN” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shai Ben-Assayag, Ran El-Yaniv
ScalableAlphaZero (SAZ) combines AlphaZero with Graph Neural Networks to enable efficient generalization of game strategies from smaller to larger board sizes. The model achieved an order of magnitude reduction in training time while maintaining or surpassing the performance of traditional AlphaZero on unseen large boards.
Key insights
- Graph Neural Networks can effectively integrate into deep reinforcement learning frameworks like AlphaZero, enabling strategy generalization across varying board dimensions without prior domain knowledge.
- Learning on small board game instances with a scalable GNN architecture can reduce overall training time by an order of magnitude compared to training directly on large boards.
- Subgraph sampling enhances MCTS decision-making by mitigating GNN prediction uncertainty, contributing to improved performance on larger, unseen board sizes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.