Re: [Computer-go] Aya's selfplay games for training value network

[Álvaro Begué]

Oh, that's unfortunate. When I read the description of what you did, I
thought the whole point of using of a sigmoid in the last stage was
precisely so the ownership map would be reasonably correct.

Your database should still be very useful, but it feels like it should be
much easier for the DCNN to learn about life and death if you can look at
who eventually owned this part of the board.

Regards,
Álvaro.




On Wed, Apr 5, 2017 at 3:51 PM, Hiroshi Yamashita <y...@bd.mbn.or.jp> wrote:

> Dear Álvaro,
>
> It looks like "remove all dead stones" doesn't always
>>
>
> I checked this. White never pass if position is losing in Tromp/Taylor
> rule. In this case, white can win even if not removing last black
> stone. And with something unlucky, pass was selected.
> And when a player is winning with big margin like +30.5, both player
> tend to play bad move, like self-atari, filling eye, breaking seki,
> if those moves do not change the result.
> So it is maybe not good at ownership map.
>
> Thanks,
> Hiroshi Yamashita
>
>
> - Original Message - From: "Álvaro Begué" <alvaro.be...@gmail.com>
> To: "computer-go" <computer-go@computer-go.org>
> Sent: Thursday, April 06, 2017 1:53 AM
> Subject: Re: [Computer-go] Aya's selfplay games for training value network
>
>
> Dear Yamashita,
>
> This is a great resource and I am very thankful that you made it available
> to all of us.
>
> One minor issue: It looks like "remove all dead stones" doesn't always
> work. An example is 206_19_0114_2k_r16_add300_7/20170312_0018_05484.sgf ,
> where a black stone at P18 is left on the board for no apparent reason. Do
> you know why white passed on move 420?
>
> I haven't studied how often this happens, but it means I will need to do
> some additional processing before I can use the final ownership map as
> something that a DCNN would predict.
>
> Thanks!
> Álvaro.
>
> ___
> 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

Re: [Computer-go] Aya's selfplay games for training value network

[Lukas van de Wiel]

Hi Yamashita

Thanks you for uploading them, and explaining how they are organised.
Your contribution is much appreciated!

Lukas

On 4/5/17, Hiroshi Yamashita  wrote:
> Hi,
>
> I have uploaded Aya's selfplay games.
> There are about 1,440,000 games.
> Selfplay program strength is about KGS 4d and 5d.
>
> Aya's selfplay games for training value network
> http://www.yss-aya.com/ayaself/ayaself.html
>
> Using these games, KGS 4d over, tygem 9d, and GoGoD, I made slightly
>  (+70 Elo) better value network than using KGS 4d over, tygem 9d,
>  and GoGoD. White tends to win in KGS komi 0.5 games. So I deleted
>  41% white win games in komi 0.5 game.
>
> Value network is 64 Filters, 14 Layers with batch normalization.
> Caffe's prototxt is
> http://www.yss-aya.com/ayaself/20170405aya_value.tar.gz
>
> A game has 4 parts.
>
> 1. First move is selected from opening book. And rorated 8 symmetry
> randomly.
> 2. 2nd - 16th moves are selected randomly from Policy network probability.
> 3. 2000 playouts/move self play. And it ends when a player resigns.
> 4. resume with 300 playouts/move selfplay from resign position. And remove
> all dead stones. This 300 playouts/move selfplay is about KGS 3d.
> Only root node is created by Policy network.
>
> Komi is 7.5, Chinese rule.
>
> Note:
> If Aya plays (1,1) or (1,19) or (19,1) or (19,19) or (9,1)(=J19) within
>  16 moves, it is bug. It should be deleted.
> If original game result "GN[B+R," and game result "RE[B+3.5]" is different
>  (like RE[W+5.5]), it should be deleted. 300 playout/move makes mistakes.
>
> Thanks,
> Hiroshi Yamashita
>
> ___
> 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