Two small neural networks started out rated below random play. They trained by self-play:
30 million games of draughts in 10 hours on one
desktop computer, and a new version was kept only when it beat the one before it. Below is the
recorded history of that run, and the networks themselves, running in your browser.
Recorded training history
How the networks got better
Run v1-2026-09-17, started 2026-09-17, commit 0a5c5a0.
Every generation then played separately seeded games against fixed opponents: random moves, a
greedy player, and minimax searches of depth 1 to 6, 134,800 arena games in all.
Ratings are Elo on that ladder, with random moves at 0.
Self-play games
30,015,856
Training time
10 hours
Arena games
134,800
Each marker is a promoted generation. Circles and the solid line: the red
network (96 hidden units).
Squares and the dashed line: the black network
(64 hidden units). Dotted
lines are the fixed opponents. Hover a marker for its interval.
The black network first played roughly even with a 4-ply
search at generation 18, after 2,656,157 self-play
games and 51 minutes of training. It scored
51.9% (95% interval 47.0% to 56.7%), which is not clearly better than an even score.
Final generation of each network: Elo with its 95% interval, and its score against each
search depth (a draw counts as half).
Network
Generation
Elo
1-ply
2-ply
3-ply
4-ply
5-ply
6-ply
red (96 hidden units)
17
1,222 (1,196 to 1,249)
100.0%
85.4%
75.5%
57.3%
55.1%
43.5%
black (64 hidden units)
22
1,191 (1,167 to 1,219)
100.0%
84.1%
70.9%
56.5%
51.4%
38.5%
Hardware: one desktop computer. The networks are small, so self-play and search ran on its
processor, and its graphics card sat mostly idle.
Live in your browser
Watch them, play them, train your own
Loading the interactive lab...
Everything in this section runs on your device, in a background worker. The page loads a small
script and a list of the networks up front. The networks themselves download only when you press
Start, and each file is checked against a published fingerprint before use.
Watch a match
Play a network
Train your own
A fresh network with random weights plays itself on your device, then plays a simple greedy opponent
before and after training so you can compare. The result is your browser's, from this one run: it
depends on how many games your device fits in the time.
This keeps one processor core busy for about 60 seconds. On a
phone, plugged in is better.
The live lab could not start in this browser. The recorded history above is unaffected.