Santorini Multi-Agent RL

Santorini is a two-player abstract strategy game played on a 5×5 grid: each turn you move a builder and construct a level of a tower, and the first player to step onto a third-level tower wins. The rules are simple, but the game tree is deep — which makes it a great testbed for reinforcement learning.

The AI is trained AlphaZero-style in PyTorch: a residual convolutional network learns policy and value purely from self-play games, guided by Monte Carlo tree search, with new networks gated against the current best before being accepted. The trained network is exported to ONNX and runs fully client-side with ONNX Runtime Web — the game engine and MCTS are ported to TypeScript, so the demo needs no server and searches happen right in your browser.

Play against it in your browser, or browse the source code on GitHub for the game engine, environment, and training code.