github lightvector/KataGo v1.13.2-kata9x9
Finetuned 9x9 Neural Net

latest releases: v1.15.1, v1.15.0, v1.14.1...
pre-release9 months ago

Marking and leaving this as a 'prerelease' since this is NOT intended to be release of a new version of KataGo's source code or binaries, but is a release of a new neural net for KataGo!
For the existing binaries and/or code version, see v1.13.2: https://github.com/lightvector/KataGo/releases/tag/v1.13.2

This is a release of a neural net specially trained for 9x9! On 9x9 boards specifically, this neural net is overall much stronger than KataGo's main distributed training nets on katagotraining.org.

Training

It was finetuned from KataGo's main run nets by data generated from 3 strong personal GPUs for several months on 9x9 games starting in many diverse positions. A large number of 9x9 positions were sampled from various datasets to provide these starting positions:

  • The move tree of an earlier failed attempt at generating a 9x9 opening book earlier this year that despite not having good evaluations, extensively covered a wide variety of 9x9 openings.
  • Manually-identified blind spot positions.
  • Top-level bot games from CGOS.
  • Human professional and amateur game collections on 9x9.
  • Collections of match games won or lost (but not drawn) by KataGo on 9x9 against other versions of itself, to focus more learning on decisive positions.
  • 9x9 games played between versions of KataGo where one side was heavily computationally advantaged against the other.
  • Various manually specified openings and handicap stone patterns.
  • A tiny number of 7x7 through 11x11 games so that the net didn't entirely forget a basic sense of scaling between board sizes.

Otherwise, the training proceeded mostly the same as KataGo's main run, with essentially the same settings.

Strength

In some fast-game tests with a few hundred playouts per move, this net has sometimes rated as much as 200 Elo stronger on 9x9 boards than KataGo's main run neural nets when using a diverse set of forced openings.

However, it's not easy to be precise because the exact amount can depend a lot on settings and the particular forced opening mixture used for games. For any particular opening or family of openings on 9x9, at top levels you can often get major nonlinearities or nontransitivities between various bots depending on what openings they just so happen to play optimally or not. This is especially the case when having bots just play from the empty board position rather than using a predetermined book or opening mixture, since the games will often severely lack opening diversity.

Also, for 9x9 because bots are strong enough that the game is highly drawish (at fair komi), the Elo difference can depend heavily on the number of visits used, as both sides approach optimal with more visits and draw increasingly frequently with fewer decisive matches.

Overall though, the net generally seems more accurate and efficient at judging common 9x9 positions.

Other Notes and Caveats

  • Don't use this net on board sizes other than 9x9! At least, don't do so while expecting it to be good, you could still do so for fun. It will in fact run on 19x19, but its evaluation quality on 19x19 has degraded in quality and drifted to be offset from fair a lot due to having months of training forgetting about 19x19 and repurposing its capacity for 9x9. It also seems to have forgotten some important joseki lines. It probably will have gotten worse at large-scale fights or big dragons as well.

  • Since it is a different net with randomly different blind spots and quirks, even on size 9x9, this finetuned net probably also has a small proportion of variations that it evaluates or plays worse than KataGo's main run nets. On average it should be much better, but of course it will not always be better.

  • One fun feature is that this net also has a little bit of training for 9x9 handicap games, including the "game" where white has a 78.5 komi** while black has 4 or 5 handicap stones, such that white wins if they live basically anywhere. This training did not reach convergence, but enough that if you try searching with a few millions of playouts, the results are pretty suggestive that white can live if black starts with all four 3-3 points, but not if black gets a fifth stone anywhere reasonable.

(**Area scoring, with 0 bonus for handicap stones as in New Zealand or Tromp-Taylor rules. If you use Chinese rules, you'll need a lower komi due to the extra compensation of N points for N handicap stones, and if you use Japanese rules you'll need a lower komi since the black stones themselves occupy space and reduce the territory. Also leaving a buffer of a few points from 9x9 = 81, like choosing 78.5 instead of 80 or 80.5 is a good idea so that the net is solidly in "if I make a living group I win" and is well separated away from "actually I always win even if I lose the whole board")

Don't miss a new KataGo release

NewReleases is sending notifications on new releases.