@Hiroshi Yamashita - about rn - cool! Really neat to see people picking up some of KataGo's methods and running them. Hopefully people will find ways to improve them further.
------ Also, in case people weren't aware since I hadn't advertised it on this list, KataGo has also been available for a while too! And has recently just had its 1.4.2 release! https://github.com/lightvector/KataGo/releases At this moment, KataGo may be the overall-strongest open source bot on all of 9x9, 13x13, and 19x19. Or I hope if not literally the strongest, then very closely in the running. * For 9x9, the version that topped http://www.yss-aya.com/cgos/9x9/standings.html not too long ago is simply just KataGo's v1.3.5 release (now superseded by v1.4.2) and using the 40 block neural net from here: https://github.com/lightvector/KataGo/releases/tag/v1.4.0. * For 19x19, the same neural net plays at a very high level too, generally stronger than Leela Zero (except for mi yuting's flying dagger). It varies by hardware and different time settings and configurations, but somewhere from 100-250 Elo would probably be typical to observe in a given test[1]. * KataGo also plays all intermediate sizes - 13, 15, even stuff like 12 - at just as high of a level. Or it should, since it trains on them all the same way. But it doesn't seem like there's notable competition going on on those sizes, even on 13x13. [1]Funnily enough, the best 19x19 tests have been with KataGo's 20 block network, *not* the 40 block network - the 40 block network is stronger per-playout, but the 20 block (which has been learning from the 40 block's games) is faster by enough to more than make up for it. Although actually the 40 block network has recently made large gains and might have caught up at time parity on 19x19 finally. There hasn't been enough testing to know for sure yet. On Sat, May 16, 2020 at 9:43 AM Hiroshi Yamashita <y...@bd.mbn.or.jp> wrote: > Hi, > > Kensuke Matsuzaki released rn-6.3.0. > It is one of the strongest 9x9 engine. > It is "LeelaZero + 9x9 + heuristic features + adjustable komi + KataGo > like learning". > > rn-6.3.0 > https://github.com/zakki/leela-zero/releases/tag/rn-6.3.0 > Rn.6.3 > https://twitter.com/k_matsuzaki/status/1260908554359173120 > > Author says v995 is a latest model, but v945 is stronger. > And v995 tends to play (4,4) on initial position. > > Thanks, > Hiroshi Yamashita > > > This is quote from README.md in zip file. > ----------------------------------------------------------------------- > # '9x9-endstate' branch > > * For 9x9 game. > * Ladder detection (by https://github.com/yssaya/leela-zero-ladder) > * Various komi (by https://github.com/ihavnoid/leela-zero) > * Additional input features. > > ----------------------------------------------------------------------- > > # 'Endstate' branch > > This is a fork of Leela Zero with the 'endstate' head. The 'stock' Leela > Zero uses the value and policy nets, while this also > predicts how the game ends. To do so, there are some changes: > > * Additional 'endstate' head : The 'endstate' is how the game ended - that > information is also stored on the training data. > * Acceleration mode : To predict the endstate, we can't just resign when > we find the game hopeless - we have to play it to the end. Hence, > once we hit the resignation threshold, we reduce the playouts to 1 > instead of resigning. > * Using the 'endstate' information as the auxillary policy - see > Network.cpp for details on how it uses the auxilary policy > > The main goal of this branch is to play reasonable handicap games (and to > some extent, play games with komi) > ----------------------------------------------------------------------- > _______________________________________________ > Computer-go mailing list > Computer-go@computer-go.org > http://computer-go.org/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
