Dominik Meier

▶ Play now

The game Quoridor fascinated me. Back in 2017 I programmed a NegaMax implementation in C++ that could beat me. Recently, with the rise of vibe coding, I picked up the project to see if I could build an AlphaZero-style bot trained by self-play. The final model beat every bot I found online, including SigmaQuoridor under different evaluation budgets. The final model was trained for 13 hours on a single Colab L4 instance. It is a ResNet with 128 channels and 10 residual blocks (12.8 MB). Code and training scripts are on GitHub.

I therefore claim to the best of my knowledge that I trained the world’s best Quoridor bot. Leave me a note if you think otherwise; I am excited to see it play against other bots.

The bot is deployed, and you can play at quoridor.dominikmeier.io.

Thanks Solving Quoridor and SigmaQuoridor for the inspiration.