Re: [computer-go] Leela experiment with 6x6
I checked these variation myself and with valkyria and it seems to be sensible. -Magnus Quoting Don Dailey <[EMAIL PROTECTED]>: I took all the games played by Leela in my 6x6 experiment and applied mini-max to them, taking the statistics at various depths into the games. Don't know if there are bugs here or not but this is what I get: Depth Score Principal Variation - --- -- 1 0.2372 C3 2 0.2372 C3 D4 3 0.2372 C3 D4 C4 4 0.2372 C3 D4 C4 D3 5 0.2500 C3 D4 C4 D3 C5 6 0.2500 C3 D4 C4 D3 C5 C2 7 0.2500 C3 D4 C4 D3 C5 C2 B2 8 0.2500 C3 D4 C4 D3 C5 C2 B2 D2 9 0.2500 C3 D4 C4 D3 C5 C2 B2 D2 E5 10 0.2500 C3 D4 C4 D3 C5 C2 B2 D2 E5 D5 11 0.2500 C3 D4 C4 D3 C5 C2 B2 D2 E5 D5 D6 12 0.2182 C3 D4 C4 D3 D5 E5 D2 E2 E3 E4 E1 F3 13 0.4286 C3 D4 C4 D3 D5 E5 D2 E2 E3 E4 E1 C5 B5 14 0.4286 C3 D4 C4 D3 D5 E5 D2 E2 E3 E4 E1 C5 B5 F3 15 0.6579 C3 D4 C4 D3 D5 E5 D2 E2 E4 E3 E6 F4 C5 C2 B2 16 0.6579 C3 D4 C4 D3 D5 E5 D2 E2 E4 E3 E6 F4 C5 C2 B2 D1 17 0.6579 C3 D4 C4 D3 D5 E5 D2 E2 E4 E3 E6 F4 C5 C2 B2 D1 B1 18 0.6579 C3 D4 C4 D3 D5 E5 D2 E2 E4 E3 E6 F4 C5 C2 B2 D1 B1 F6 19 0.7000 C3 D4 C4 D3 D2 E2 D5 E5 E3 E4 E1 C5 B5 F3 C2 D6 B6 F1 F2 20 0.7000 C3 D4 C4 D3 D2 E2 D5 E5 E3 E4 E1 C5 B5 F3 C2 D6 B6 F1 F2 C6 21 1. C3 D4 C4 D3 D2 C5 D5 E5 E2 B5 B4 F3 F2 A4 ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Leela experiment with 6x6
Thanks. - Don On Fri, 2008-10-03 at 11:56 +0200, Magnus Persson wrote: > I checked these variation myself and with valkyria and it seems to be > sensible. > > -Magnus > > Quoting Don Dailey <[EMAIL PROTECTED]>: > > > I took all the games played by Leela in my 6x6 experiment and applied > > mini-max to them, taking the statistics at various depths into the > > games. > > > > Don't know if there are bugs here or not but this is what I get: > > > > Depth Score Principal Variation > > - --- -- > > 1 0.2372 C3 > > 2 0.2372 C3 D4 > > 3 0.2372 C3 D4 C4 > > 4 0.2372 C3 D4 C4 D3 > > 5 0.2500 C3 D4 C4 D3 C5 > > 6 0.2500 C3 D4 C4 D3 C5 C2 > > 7 0.2500 C3 D4 C4 D3 C5 C2 B2 > > 8 0.2500 C3 D4 C4 D3 C5 C2 B2 D2 > > 9 0.2500 C3 D4 C4 D3 C5 C2 B2 D2 E5 > > 10 0.2500 C3 D4 C4 D3 C5 C2 B2 D2 E5 D5 > > 11 0.2500 C3 D4 C4 D3 C5 C2 B2 D2 E5 D5 D6 > > 12 0.2182 C3 D4 C4 D3 D5 E5 D2 E2 E3 E4 E1 F3 > > 13 0.4286 C3 D4 C4 D3 D5 E5 D2 E2 E3 E4 E1 C5 B5 > > 14 0.4286 C3 D4 C4 D3 D5 E5 D2 E2 E3 E4 E1 C5 B5 F3 > > 15 0.6579 C3 D4 C4 D3 D5 E5 D2 E2 E4 E3 E6 F4 C5 C2 B2 > > 16 0.6579 C3 D4 C4 D3 D5 E5 D2 E2 E4 E3 E6 F4 C5 C2 B2 D1 > > 17 0.6579 C3 D4 C4 D3 D5 E5 D2 E2 E4 E3 E6 F4 C5 C2 B2 D1 B1 > > 18 0.6579 C3 D4 C4 D3 D5 E5 D2 E2 E4 E3 E6 F4 C5 C2 B2 D1 B1 F6 > > 19 0.7000 C3 D4 C4 D3 D2 E2 D5 E5 E3 E4 E1 C5 B5 F3 C2 D6 B6 F1 F2 > > 20 0.7000 C3 D4 C4 D3 D2 E2 D5 E5 E3 E4 E1 C5 B5 F3 C2 D6 B6 F1 F2 > > C6 > > 21 1. C3 D4 C4 D3 D2 C5 D5 E5 E2 B5 B4 F3 F2 A4 > > > > > > > ___ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ signature.asc Description: This is a digitally signed message part ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] Congratulations to David Fotland!
Currently Many Faces' strength peaks at 32 cores, and goes down with more. On a single core, strength increases 50 to 80 ELO points per doubling of performance. On multiple processors the rate of increase is much less. The increase is very sensitive to communication latency, so a [EMAIL PROTECTED] type of engine is unlikely to be much stronger than a single computer or local cluster. The 32 core machine I'm using has 40 Gbps Infiniband interconnect with less than 2 microsecond latency. David -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of Michael Markefka Sent: Thursday, October 02, 2008 10:18 AM To: computer-go Subject: Re: [computer-go] Congratulations to David Fotland! So, when are we going to see distributed computing? [EMAIL PROTECTED], [EMAIL PROTECTED], [EMAIL PROTECTED] With Go engines that scale well to increased processing capacity, imagine facilitating a few thousand PCs to do the computing. For good measure, [EMAIL PROTECTED] as about 800,000 nodes online as of now. What's the approximate increase in playing level per increase in processing power? Any rough law for that? Best regards, Mike Olivier Teytaud wrote: > Mogo was allowed to use 800 cores, not more, and only for games against > humans. > We have no acces to so many cores for computer-computer games (if there > were only three teams involved, > we could :-) ). > For some games Huygens was unaivalable at all, and mogo played with much > weaker hardware (some quad-cores, > however, it is not so bad :-) ). > > Best regards, > Olivier > > > > > ___ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] 7.5-komi for 9x9 in Beijing
It seems likely now that the correct komi for 9x9 is 7.0. If so, I'd prefer 6.5 komi to 7.5, since 6.5 would have black winning most games, and most other games have a first player advantage. This would give 9x9 go a similar first player advantage. David -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of "Ingo Althöfer" Sent: Thursday, October 02, 2008 10:35 AM To: computer-go@computer-go.org Subject: [computer-go] 7.5-komi for 9x9 in Beijing Gian-Carlo Pascutto wrote: > I'd have some preference for playing the decisive game > with komi = 6.5, but apparently thats not possible on KGS. But that should not be a problem, as long as the operators do not believe in the final verdict of KGS. > I think with komi = 7.5 white > is scoring very high (too high?) in the top games. I made a count for the 9x9-competition in Beijing (komi=7.5: Looking only at games among the top 5 rankers there are 20 games so far (including two tiebreak-games) with 15 wins for White and 5 Wins for Black. Looking at all games among the top 7 rankers there are 40 games (including two tiebrak-games) with 27 : 13 for White. Taking all 164 games of the tournament (including two tiebrak games) White is ahead by 93:71. It is probably not an accident that the quota decrease with decreasing playing strength: 15/20 > 27/40 > 93/164 or translated to decimals 0.75 > 0.675 > 0.567. However, when B+7 is the correct value on 9x9, komi=6.5 might lead into similar problems. Ingo. -- Pt! Schon vom neuen GMX MultiMessenger gehört? Der kann`s mit allen: http://www.gmx.net/de/go/multimessenger ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] On ranks 2 and 3 of 9x9 in Beijing
No. on 9x9 without sekis, the score must be odd, 5, 7, or 9. From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of Antonin Lucas Sent: Thursday, October 02, 2008 12:16 PM To: computer-go Subject: Re: [computer-go] On ranks 2 and 3 of 9x9 in Beijing On Thu, Oct 2, 2008 at 6:13 PM, Gian-Carlo Pascutto <[EMAIL PROTECTED]> wrote: I'd have some preference for playing the decisive game with komi = 6.5, but apparently thats not possible on KGS. I think with komi = 7.5 white is scoring very high (too high?) in the top games. Aren't 6.5 and 7.5 komi in area counting essentially equivalent, save for a few seki cases ? ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] 7.5-komi for 9x9 in Beijing
I had heard somewhere that there are some who believe 8.0 is the right komi for 9x9 Chinese. I personally believed for a long time it was 7.0 based on statistical data of games.However that can be misleading. I noticed with CGOS data that White has a very slight edge with 7.5 but I also noticed by mini-maxing the tree for just the first 2 moves that black actually has a slight advantage if white plays the correct response to whatever blacks first move is. As it turns out the "correct" response is not quite as frequently played. Of course this mean little - it's just statistical data. However, I would prefer using the komi that tends to give the closest scores and I don't know if that is 6.5 or 7.5.So again, if 7.0 is correct, then (to me) the question is which komi gives the most even results. 6x6 is an interesting test case. It appears from Leela data that with 2.5 komi, black should win - but in practice black is much harder to play correctly.Maybe it's that way in 9x9? I would not be surprised if something like that was true of 9x9. So it is possible that with 7.5 komi black DOES have an advantage - but it's just more difficult to play. I'm just hypothesizing here. I guess I like 7x7 because statistically it's very very close and I suspect that if we changed to 6.5 there would be a significant bias in favor of black - but I don't know that for sure. - Don On Fri, 2008-10-03 at 05:12 -0700, David Fotland wrote: > It seems likely now that the correct komi for 9x9 is 7.0. If so, I'd prefer > 6.5 komi to 7.5, since 6.5 would have black winning most games, and most > other games have a first player advantage. This would give 9x9 go a similar > first player advantage. > > David > > -Original Message- > From: [EMAIL PROTECTED] > [mailto:[EMAIL PROTECTED] On Behalf Of "Ingo Althöfer" > Sent: Thursday, October 02, 2008 10:35 AM > To: computer-go@computer-go.org > Subject: [computer-go] 7.5-komi for 9x9 in Beijing > > Gian-Carlo Pascutto wrote: > > I'd have some preference for playing the decisive game > > with komi = 6.5, but apparently thats not possible on KGS. > > But that should not be a problem, as long as the operators > do not believe in the final verdict of KGS. > > > I think with komi = 7.5 white > > is scoring very high (too high?) in the top games. > > I made a count for the 9x9-competition in Beijing (komi=7.5: > > Looking only at games among the top 5 rankers > there are 20 games so far (including two tiebreak-games) > with 15 wins for White and 5 Wins for Black. > > Looking at all games among the top 7 rankers > there are 40 games (including two tiebrak-games) > with 27 : 13 for White. > > Taking all 164 games of the tournament (including two > tiebrak games) White is ahead by 93:71. > > It is probably not an accident that the quota decrease > with decreasing playing strength: > > 15/20 > 27/40 > 93/164 or translated to decimals > 0.75 > 0.675 > 0.567. > > However, when B+7 is the correct value on 9x9, > komi=6.5 might lead into similar problems. > > Ingo. > signature.asc Description: This is a digitally signed message part ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] On ranks 2 and 3 of 9x9 in Beijing
So are you saying that you can get even scores, but most of the games will have odd scores? So it's possible for the correct komi to be even if best play must result in a seki situation? - Don On Fri, 2008-10-03 at 05:29 -0700, David Fotland wrote: > No. on 9x9 without sekis, the score must be odd, 5, 7, or 9. > > > > From:[EMAIL PROTECTED] > [mailto:[EMAIL PROTECTED] On Behalf Of Antonin > Lucas > Sent: Thursday, October 02, 2008 12:16 PM > To: computer-go > Subject: Re: [computer-go] On ranks 2 and 3 of 9x9 in Beijing > > > > > On Thu, Oct 2, 2008 at 6:13 PM, Gian-Carlo Pascutto <[EMAIL PROTECTED]> > wrote: > > > > I'd have some preference for playing the decisive game with > komi = 6.5, > but apparently thats not possible on KGS. I think with komi = > 7.5 white > is scoring very high (too high?) in the top games. > > > Aren't 6.5 and 7.5 komi in area counting essentially equivalent, save > for a few seki cases ? > > > ___ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ signature.asc Description: This is a digitally signed message part ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] Congratulations to David Fotland!
Microsoft just released Windows HPC server 2008, a server operating system for high performance clusters. Their MPI implementation is about 10% faster than Linux on the huge machines with thousands of cores. I'm working with the Microsoft HPC group. They have been exceptionally helpful. David -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of Mark Boon Sent: Thursday, October 02, 2008 7:57 PM To: computer-go Subject: Re: [computer-go] Congratulations to David Fotland! Is Microsoft now selling computers? Interesting... Let me chime in with my congratulations to David. Mark On 2-okt-08, at 20:52, Darren Cook wrote: >> investment. If we can find corporate sponsors, it should not be hard >> to gain access to such hardware. Reading between the lines, I think >> some Microsoft wunderkind may be backing Dave Fotland. > > > It seems Microsoft are selling such hardware and approached David > while > looking for some application that could give them PR for it. (And the > new Many Faces will be bundled with it apparently!) > > I was speaking to this Microsoft Go Enthusiast yesterday and there > is a > chance of even more support from Microsoft for game AI in the future. > > > Darren > ___ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] On ranks 2 and 3 of 9x9 in Beijing
Yes. An even score requires a seki with an odd number of neutral (uncounted) points. So even komi is possible, but unlikely since very few games end with seki, unless the score is very unbalanced. David -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of Don Dailey Sent: Friday, October 03, 2008 5:37 AM To: computer-go Subject: RE: [computer-go] On ranks 2 and 3 of 9x9 in Beijing So are you saying that you can get even scores, but most of the games will have odd scores? So it's possible for the correct komi to be even if best play must result in a seki situation? - Don On Fri, 2008-10-03 at 05:29 -0700, David Fotland wrote: > No. on 9x9 without sekis, the score must be odd, 5, 7, or 9. > > > > From:[EMAIL PROTECTED] > [mailto:[EMAIL PROTECTED] On Behalf Of Antonin > Lucas > Sent: Thursday, October 02, 2008 12:16 PM > To: computer-go > Subject: Re: [computer-go] On ranks 2 and 3 of 9x9 in Beijing > > > > > On Thu, Oct 2, 2008 at 6:13 PM, Gian-Carlo Pascutto <[EMAIL PROTECTED]> > wrote: > > > > I'd have some preference for playing the decisive game with > komi = 6.5, > but apparently thats not possible on KGS. I think with komi = > 7.5 white > is scoring very high (too high?) in the top games. > > > Aren't 6.5 and 7.5 komi in area counting essentially equivalent, save > for a few seki cases ? > > > ___ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Using playouts for more than position evaluation?
>>> For instance, if an intersection belongs to the same colour in all >>> playouts, chances are >>> that it is fairly secure (that doesn't mean one >> shouldn't play there, sacrifices there may have an impact on other >> intersections). Ok, that one was well know (ownership maps, territory heuristic). >> Or, if an intersection is black in all playouts won by black, and white >> in all playouts won >> by white, chances are that it is fairly important to play there (since >> playouts are random, >> there is no guarantee, but emphasizing such intersections, and their >> ordering, in the top-level >> tree search seems profitable). That one was related to some existing work, and though perhaps not exactly the same, already under research: > We (the Orego team) have done some work along these lines this summer. We're > working on a paper. There is also this paper, right on topic: Combining tactical search and Monte-Carlo in the game of Go Tristan Cazenave, Bernard Helmstetter http://www.citeulike.org/group/5884/article/2990531 That still seems to leave some other possibilities, though. For instance, what about intersections that are 50/50, independent of who plays them first? If they exist, it would seem that their fate is determined by plays elsewhere, in particular, it doesn't seem like a good idea to play there, and if one does already have a stone there, one might have to do something about it. Also, in a balanced game, one would like to see some balance of undecided intersections. Or what about keeping the difference between statistical evaluations for a few moves? Our own moves are driven by having positive effects on those values, but if an opponent's move changes only one of the values (without offering compensation in another value), one would do well to answer it, and the particular values affected (does it unsettle a stable group, or influence previously neutral territory, or ..) might give an indication of the nature of the imbalance, and possible ways to counter it (sufficiently to restore the balance). Do these make sense? And are there other useful items one might be able to extract from the playout statistics, online, during the game (as opposed to offline learning, from self-play)? Of course, some of this could be expected to emerge from playouts without additional intervention, eg, if one always tries to optimize the statistical results, then good answers to imbalances introduced by opponent play should emerge. But given the size of the search space, having more concrete handles on information to help directing the search might still be useful. Claus ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Using playouts for more than position evaluation?
We've got some decent results (or at least interesting pictures) by looking at the correlation between controlling a particular point and winning the game. Peter Drake http://www.lclark.edu/~drake/ On Oct 3, 2008, at 7:47 AM, Claus Reinke wrote: Do these make sense? And are there other useful items one might be able to extract from the playout statistics, online, during the game (as opposed to offline learning, from self-play)? ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] Bad effect when tree node limit reached
My engine is basic UCT+MC. I've been letting it ponder on small empty boards with a certain komi to see if it begins to converge to the correct position value. It does as long as the tree is allowed to grow. But as soon as the tree reaches its node limit (with simulations continuing), the position value begins to go back toward 0.5 (where 0 represents a white win and 1 represents a black win). I would have expected the convergence to the correct value to continue at a slower rate or maybe just stay the same. Is this a bug or can this effect be explained? ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Bad effect when tree node limit reached
Hi; the "node limit" is a hard limit on the number of nodes ? if yes, when you reach this limit, what do you do - you remove some nodes with small number of simulations ? Olivier My engine is basic UCT+MC. I've been letting it ponder on small empty > boards with a certain komi to see if it begins to converge to the correct > position value. It does as long as the tree is allowed to grow. But as > soon as the tree reaches its node limit (with simulations continuing), the > position value begins to go back toward 0.5 (where 0 represents a white win > and 1 represents a black win). I would have expected the convergence to the > correct value to continue at a slower rate or maybe just stay the same. Is > this a bug or can this effect be explained? > ___ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ > -- = Olivier Teytaud (TAO-inria) [EMAIL PROTECTED] Tel (33)169154231 / Fax (33)169156586 Equipe TAO (Inria-Futurs), LRI, UMR 8623(CNRS - Universite Paris-Sud), bat 490 Universite Paris-Sud 91405 Orsay Cedex France ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
