search

lightvector / KataGo

GTP engine and self-play learning in Go
https://katagotraining.org/
Other
3.53k stars 564 forks source link

Why does this no-variation opening appear so frequently in KataGo’s training games? #993

Open poptangtwe opened 5 days ago

poptangtwe commented 5 days ago

There is a no-variation opening in training games, and it appears very very frequently. https://katagotraining.org/sgfplayer/training-games/61976462/ https://katagotraining.org/sgfplayer/training-games/60902084/ a no-variation opening in training games Why does this opening appear so frequently in KataGo’s self-play game records, with almost no variation from the 1st to the 44th move every time it occurs? It has been like this for quite a long time, maybe even earlier before 60B. Could this affect KataGo’s training? The similar issue has been around since the era of Leela 40B. Is there any way to technically reduce the occurrence frequency of the opening? For example, by increasing the self-play temperature or automatically filtering out these repetitive positions during training? And I suggest continuing to increase the diversity of openings in self-play games, exploring more possibilities, and avoiding overfitting.

poptangtwe commented 5 days ago

https://katagotraining.org/sgfplayer/training-games/63975130/ IMG_3343 This is another variant of the opening, which now appears with a very high frequency. It is almost no-change from the 1st to the 40th move.

HackYardo commented 3 days ago

Maybe it is a pretty opening? (although I don't think so) What are the most frequent openings that LeelaZero 40 series weights trained, the same as KataGo? Does this sort of behaviors stop LeelaZero from getting more powerful? From my rough view, openings could easily be changed by set kata-param wideRootNoise.

poptangtwe commented 2 days ago

Maybe it is a pretty opening? (although I don't think so) What are the most frequent openings that LeelaZero 40 series weights trained, the same as KataGo? Does this sort of behaviors stop LeelaZero from getting more powerful? From my rough view, openings could easily be changed by set kata-param wideRootNoise.

Leela’s repetitive opening moves are much more than KataGo’s.

Perhaps the noise in training games is too low? I’m worried that too many repetitive moves in the opening phase may affect the improvement of KataGo's ELO.

michito744 commented 2 days ago

@poptangtwe

KataGo's search is significantly less diverse with too little fluctuation.

Opening of the 7X7 board JP rules 2024-10-19 (3)

B(A6) on move 13 is extremely promising and appears in the opening book, but the policy weighting is less than 1% so it is completely ignored in the default setting . 2024-10-19 (2)

poptangtwe commented 18 hours ago

https://katagotraining.org/sgfplayer/training-games/64162195/ IMG_3502 This is another variant of this opening, which appears with a high frequency recently. It is almost no-change from the 1st to the 27th move.