Hi,

> using rollouts to compensate for Leela's network being trained with the
"wrong" komi for this competition:

Yes, and it seems that rollouts isn't useful when trained komi is "correct".

> Our program Natsukaze also used Leela Zero recent 700000 selfplay games
to train DNN.

What would happen if Natsukaze used filtered raw training data instead of
filtered self play games?

https://github.com/gcp/leela-zero/issues/167

2018年12月19日(水) 7:35 Hiroshi Yamashita <y...@bd.mbn.or.jp>:

> Hi,
>
> Our program Natsukaze also used Leela Zero recent 700000 selfplay games
>   to train DNN.
> Ladder escape moves(4% of total games) are removed, and chasing not
>   ladder(0.3%) also removed. But its DNN policy was weak, around CGOS 2100.
>
> Maybe it is because current LZ selfplay use t=1 not first 30 moves but all
> moves.
> I did not know this. I think this makes selfplay weaker +1000 Elo.
>
> Switch to t=1 for all self-play moves, i.e., randomcnt=999
> https://github.com/gcp/leela-zero-server/pull/81
>
> Thanks,
> Hiroshi Yamashita
>
>
> On 2018/12/19 2:01, Gian-Carlo Pascutto wrote:
> > On 17/12/18 01:53, Hiroshi Yamashita wrote:
> >> Hi,
> >>
> >> AI Ryusei 2018 was held on 15,16th December in Nihon-kiin, Japan.
> >> 14 programs played preliminary swiss 7 round, and top 6 programs
> >>   played round-robin final. Then, Golaxy won.
> >>
> >> Result
> >> https://www.igoshogi.net/ai_ryusei/01/en/result.html
> >
> > It appears the 2nd place finisher after Golaxy was a hybrid of Rn and
> > Leela Zero, using rollouts to compensate for Leela's network being
> > trained with the "wrong" komi for this competition:
> >
> > https://github.com/zakki/Ray/issues/171#issuecomment-447637052
> > https://img.igoshogi.net/ai_ryusei/01/data/11.pdf
> >
> _______________________________________________
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go

-- 
MATSUZAKI Kensuke
_______________________________________________
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Reply via email to