AlphaZero training lab

A convolutional policy+value net plays itself in your browser (in a Web Worker) and learns from the games — pure self-play, no framework. Watch the live self-play game, the training loss, and its win-rate against fixed opponents climb as it improves.

idle

Live self-play

games0

positions0

elapsed0s

games/s0.0

loss0.000

Training loss

Win-rate vs fixed opponents

vs Random vs AlphaBeta-d2 0–100%

This is a genuine (if tiny, CPU-only) AlphaZero loop: PUCT self-play → policy+value training → repeat. It learns to beat Random within a minute; beating a deep search needs far more compute.