KataGo Distributed Training

Rectangular Board Testing and Solving!

This is a site with test runs conducted in late 2020 to exercise and debug KataGo's distributed training infrastructure. The main test run "rect15" trained on board sizes ranging from 3x3 to 15x15 including all rectangular sizes.

Although just a test, the nets here are strong and should generally be superhuman. For board sizes in this range that are extremely non-square, these are probably the strongest available KataGo nets, and there is a good chance they have optimally solved all the usual smaller rectangular sizes. (But mind that the exact ruleset can matter for some weird sizes, and KataGo's formal rules may slightly differ in sufficiently weird corner cases from human rulesets.)

This run is finished now and is left here as an archive for anyone interested. You can explore it and try the nets if you are interested in what near-optimal play might be on weird board sizes like 3x9, or 4x7, or 5x6. Enjoy!

Stats for rect15

This run is named rect15 and began on 2020-09-09 04:00:46 UTC.

Data Generation

Across all time, 10 distinct users have uploaded 97,313,769 rows of training data, 4,233,252 training games, and 99,711 rating games.

Look up and view games for this run here.

Networks

A total of 651 networks have been uploaded for rect15.

Latest network: rect15-b20c256-s343365760-d96847752

Strongest confidently-rated network: rect15-b15c192-s403552640-d80841171

Approximate Elo Ratings Graph
Graph is based on about 99,711 rating games using mid to high hundreds of playouts. Ratings might still be mildly inflated due to only playing other KataGo nets, but otherwise are fresh and unbiased and involve a variety of nets to avoid rock-paper-scissors. Vertical bars indicate approximately a 95% confidence interval.

Graph is based on games at equal visits rather than equal compute time, since the former is easier to control on widely varying user hardware. Larger/slower networks will be advantaged due to getting more total compute time. Given particular hardware, one can approximately correct for this by benchmarking to measure the speed difference, each doubling of speed may be worth somewhere around 150-250 Elo.

Click and drag to zoom. Double-click or click on a button to reset zoom.

Approx Elo given equal time rather than equal visits (might NOT be accurate, varies with hardware)

Top Contributors

See here for a full list of contributors for rect15.

Overall

User Training Rows Training Games Rating Games
lightvector 47,718,820 rows generated 2,093,875 training games 52,420 rating games
petgo3 24,259,887 rows generated 1,056,724 training games 23,250 rating games
JBX2010 16,449,638 rows generated 697,024 training games 13,482 rating games
iopq 8,660,205 rows generated 375,894 training games 10,374 rating games
sbbdms 171,643 rows generated 7,276 training games 112 rating games
nebu 43,973 rows generated 1,969 training games 72 rating games
luckyangman 4,654 rows generated 204 training games 0 rating games
sander 3,930 rows generated 230 training games 1 rating games
ccytest 782 rows generated 43 training games 0 rating games
cheeseonhead 237 rows generated 13 training games 0 rating games

Older Runs

For older or prior runs that KataGo has done on this site, see here.