I built Miai, a full-contract Bridge bot trained from scratch, and want to hear your feedback
Hello everyone,
I'm Zhiyuan, a Ph.D. student at MIT's Game Solving Lab. I've been developing a new Bridge bot called Miai, which learns Bridge from scratch without human gameplay data. To the best of my knowledge, this is the first full contract Bridge agent trained this way, covering both bidding and card play.
I made a small website where you can try it here:
One thing I found especially interesting is that Miai seems to have developed a mostly natural and interpretable bidding system. Its 1NT opening is natural, balanced or semi-balanced, roughly 13-17 HCP. Its two-level suit openings, including 2♣, are weak natural openings. Responses are natural by default, and many game/slam decisions are direct placements. I have not found clear transfer or relay structures so far.
I wrote up my current analysis of its system here:
https://bridge.miai.moe/systemd
I am not sure how strong Miai is against top human players, but it does beat the bots I've been able to test against. The version currently on the website is a weaker version that acts instantly due to computational constraints. In a 1,024-board duplicate evaluation, this version already scored +0.21 IMP/board against WBridge5, while the contracts it reached scored about +0.59 IMP/board by double-dummy analysis. So I suppose it is playing something meaningful.
I'd love to share this with the Bridge community and hear your feedback. Please feel free to try playing against Miai, inspect the system, and let me know what looks strong, weak, weird, illegal, or just interesting.
Bug reports, strange auctions, and general feedback are very welcome. You can also contact me at:
fanzy [at] mit [dot] edu
Thanks!
