Re: [computer-go] Black/White winning rates with random playout?
Nick Wedd wrote: I suggest that instead of getting your neural players to play Go, you get them to play a very slightly different game, in which, when both players pass in turn, all stones remaining on the board are deemed alive. It is not difficult to write a scoring algorithm for this game. I believe this game is called Go with Tromp-Taylor rules. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Black/White winning rates with random playout?
On Sun, Jan 11, 2009 at 01:24:32PM +, Nick Wedd wrote: I suggest that instead of getting your neural players to play Go, you get them to play a very slightly different game, in which, when both players pass in turn, all stones remaining on the board are deemed alive. It is not difficult to write a scoring algorithm for this game. Or you can rephrase this to say that your neural players should play Go using the Tromp-Taylor ruleset. Scoring is pretty much trivial to implement in these rules, and they approximate the traditional chinese counting relatively well - all my bots always played on KGS using just the Tromp-Taylor counting and discrepances are rare. -- Petr Pasky Baudis The average, healthy, well-adjusted adult gets up at seven-thirty in the morning feeling just terrible. -- Jean Kerr ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] Black/White winning rates with random playout?
Just a remark, no advice: the suicide rule, the ko rule and the score depend on the rules under which the game is played. Dave Van: computer-go-boun...@computer-go.org namens Ernest Galbrun Verzonden: za 10-1-2009 22:01 Aan: computer-go Onderwerp: Re: [computer-go] Black/White winning rates with random playout? Hello everyone, I am trying to do some genetic experiment with virtual go players I programmed using basic neuronal network technology. The principle is to test my randomly mutated players against each other and to kill the losers. I have used Opengo library to make my players play against each other, the problem is that opengo does not have any scoring capability, so I am never certain about the result of a game ; and opengo has still some bugs, especially when making computer players play against each other. Could someone tell me what program (if any) I could use to make the players play against each other, given that : - My players dont know anything about the suicide and the ko rule. - Thay can't count the score by themselves. Ernest Galbrun ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Black/White winning rates with random playout?
In message 87bb3fc30901101301tad289d8g1d7cd7e14b731...@mail.gmail.com, Ernest Galbrun ernest.galb...@gmail.com writes Hello everyone, I am trying to do some genetic experiment with virtual go players I programmed using basic neuronal network technology. The principle is to test my randomly mutated players against each other and to kill the losers. I have used Opengo library to make my players play against each other, the problem is that opengo does not have any scoring capability, so I am never certain about the result of a game ; and opengo has still some bugs, especially when making computer players play against each other. Could someone tell me what program (if any) I could use to make the players play against each other, given that : - My players dont know anything about the suicide and the ko rule. - Thay can't count the score by themselves. I am not sure what you mean by the result of a game. If a game has stopped because two weak players have passed in turn, then the result may, depending on the rules used, be undefined, or difficult or inappropriate to calculate. If a game has stopped because two expert players have passed in turn, then the result is (almost certainly) defined, but it may be difficult for medium-strength players, and impossible for all existing programs, to know what it is. I suggest that instead of getting your neural players to play Go, you get them to play a very slightly different game, in which, when both players pass in turn, all stones remaining on the board are deemed alive. It is not difficult to write a scoring algorithm for this game. Nick -- Nick Weddn...@maproom.co.uk ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Black/White winning rates with random playout?
Le jeudi 8 janvier 2009, Nuno Milheiro a écrit : It seems normal to me, Blac is only one play ahead, which value is several points (probably 7,5 hence the komi value) given intelligent play, given random play the value of one more move may be only one point. You should try with more komi value to see which is the fair komi value for random play 2009/1/8 i...@gmx.ch Hi, What's an usual winning rate for black/white from an empty 9x9 board, black playing first, 7.5 komi? I play 50k games when starting my program, and I usually get around 60% winning rate for white. This seems rather high to me, and I suspect a bug somewhere. Do you have any data I could compare with? I use random playouts, only moves that fill eyes are pruned. This has been discussed some times ago wrt correct komi. Aloril did some experiments with various programs, and this tends to confirm that the komi depends on the strenght of the player. For a random player, komi is near 1 point = mostly the advantage of plonking one more stone on the board. Alain ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Black/White winning rates with random playout?
Hello everyone, I am trying to do some genetic experiment with virtual go players I programmed using basic neuronal network technology. The principle is to test my randomly mutated players against each other and to kill the losers. I have used Opengo library to make my players play against each other, the problem is that opengo does not have any scoring capability, so I am never certain about the result of a game ; and opengo has still some bugs, especially when making computer players play against each other. Could someone tell me what program (if any) I could use to make the players play against each other, given that : - My players dont know anything about the suicide and the ko rule. - Thay can't count the score by themselves. Ernest Galbrun ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] Black/White winning rates with random playout?
Hi, What's an usual winning rate for black/white from an empty 9x9 board, black playing first, 7.5 komi? I play 50k games when starting my program, and I usually get around 60% winning rate for white. This seems rather high to me, and I suspect a bug somewhere. Do you have any data I could compare with? I use random playouts, only moves that fill eyes are pruned. Cheers, ibd -- 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/
Re: [computer-go] Black/White winning rates with random playout?
On Thu, 2009-01-08 at 19:27 +0100, i...@gmx.ch wrote: Hi, What's an usual winning rate for black/white from an empty 9x9 board, black playing first, 7.5 komi? I play 50k games when starting my program, and I usually get around 60% winning rate for white. This seems rather high to me, and I suspect a bug somewhere. Do you have any data I could compare with? I use random playouts, only moves that fill eyes are pruned. Cheers, ibd I don't know the answer, but it's not too surprising - with random play the komi should be something like 2 or 3, so white with 7.5 komi has a pretty good advantage. This advantage disappears (or almost disappears) if the games are well played, but in your case they are not. - Don ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Black/White winning rates with random playout?
It seems normal to me, Blac is only one play ahead, which value is several points (probably 7,5 hence the komi value) given intelligent play, given random play the value of one more move may be only one point. You should try with more komi value to see which is the fair komi value for random play 2009/1/8 i...@gmx.ch Hi, What's an usual winning rate for black/white from an empty 9x9 board, black playing first, 7.5 komi? I play 50k games when starting my program, and I usually get around 60% winning rate for white. This seems rather high to me, and I suspect a bug somewhere. Do you have any data I could compare with? I use random playouts, only moves that fill eyes are pruned. Cheers, ibd -- 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/ -- Nuno Milheiro -BEGIN GEEK CODE BLOCK- Version: 3.12 GCM d+! s: a C ULAS$ P+ L+++$ E+++ W+ N o? K- w--$ O? M- V-- PS+++ PE- Y+ PGP++ t 5? X+ !R tv- b+ DI++ D+ G+ e+++ h r+++ y* --END GEEK CODE BLOCK-- ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Black/White winning rates with random playout?
Sounds about right. Looking at my notes, I have 57% wins for white using similar playout?rules. - Dave Hillis -Original Message- From: Don Dailey dailey@gmail.com To: computer-go computer-go@computer-go.org Sent: Thu, 8 Jan 2009 1:38 pm Subject: Re: [computer-go] Black/White winning rates with random playout? On Thu, 2009-01-08 at 19:27 +0100, i...@gmx.ch wrote: Hi, What's an usual winning rate for black/white from an empty 9x9 board, black playing first, 7.5 komi? I play 50k games when starting my program, and I usually get around 60% winning rate for white. This seems rather high to me, and I suspect a bug somewhere. Do you have any data I could compare with? I use random playouts, only moves that fill eyes are pruned. Cheers, ibd I don't know the answer, but it's not too surprising - with random play the komi should be something like 2 or 3, so white with 7.5 komi has a pretty good advantage. This advantage disappears (or almost disappears) if the games are well played, but in your case they are not. - Don ___ 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] Black/White winning rates with random playout?
I ran some tests with 500k games each and came to this result: with komi 0.5, white has 47.5 winn. perc. with komi 1.5, white has 50.7 winn. perc. with komi 2.5, white has 50.9 winn. perc. with komi 3.5, white has 54.0 winn. perc. with komi 4.5, white has 53.8 winn. perc. -- ? with komi 5.5, white has 57.3 winn. perc. with komi 6.5, white has 57.1 winn. perc. - ? with komi 7.5, white has 60.3 winn. perc. with komi 8.5, white has 60.3 winn. perc. with komi 9.5, white has 63.1 winn. perc. with komi 10.5, white has 63.1 winn. perc. with komi 11.5, white has 65.9 winn. perc. The percentage is mostly according to komi, with 2 exceptions. -- 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/
Re: [computer-go] Black/White winning rates with random playout?
You won't get any playouts whose outcome is even, so 3.5 and 4.5 are effectively the same komi in this experiment (it would be different if seki were possible, but naive playouts don't result in seki). Your results seem very plausible to me. Álvaro. On Thu, Jan 8, 2009 at 1:59 PM, Isaac Deutsch i...@gmx.ch wrote: I ran some tests with 500k games each and came to this result: with komi 0.5, white has 47.5 winn. perc. with komi 1.5, white has 50.7 winn. perc. with komi 2.5, white has 50.9 winn. perc. with komi 3.5, white has 54.0 winn. perc. with komi 4.5, white has 53.8 winn. perc. -- ? with komi 5.5, white has 57.3 winn. perc. with komi 6.5, white has 57.1 winn. perc. - ? with komi 7.5, white has 60.3 winn. perc. with komi 8.5, white has 60.3 winn. perc. with komi 9.5, white has 63.1 winn. perc. with komi 10.5, white has 63.1 winn. perc. with komi 11.5, white has 65.9 winn. perc. The percentage is mostly according to komi, with 2 exceptions. -- 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] Black/White winning rates with random playout?
Isaac Deutsch wrote: I ran some tests with 500k games each and came to this result: with komi 0.5, white has 47.5 winn. perc. with komi 1.5, white has 50.7 winn. perc. with komi 2.5, white has 50.9 winn. perc. with komi 3.5, white has 54.0 winn. perc. with komi 4.5, white has 53.8 winn. perc. -- ? with komi 5.5, white has 57.3 winn. perc. with komi 6.5, white has 57.1 winn. perc. - ? with komi 7.5, white has 60.3 winn. perc. with komi 8.5, white has 60.3 winn. perc. with komi 9.5, white has 63.1 winn. perc. with komi 10.5, white has 63.1 winn. perc. with komi 11.5, white has 65.9 winn. perc. The percentage is mostly according to komi, with 2 exceptions. If your playouts don't score seki, then komi = 2n + 1.5 and komi = 2n + 2.5 are equivalent. So you should get the same winn. perc. for 1.5 and 2.5. Ditto for 3.5 and 4.5, etc. That's because White Score + Black Score = 81, so Black Score - WhiteScore = 81 - 2 * WhiteScore is always odd. Rémi ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Black/White winning rates with random playout?
I don't know the answer, but it's not too surprising - with random play the komi should be something like 2 or 3, so white with 7.5 komi has a pretty good advantage. This advantage disappears (or almost disappears) if the games are well played, but in your case they are not. I think that the advantage increases when the games are well played, and I believe some people with other programs have the same results. In particular with good openings, white becomes very strong. Also, high level human players also have this feeling (I think). Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Black/White winning rates with random playout?
Nuno Milheiro wrote: Given random play, komi value does not change play, so we could see what is the mean score (no komi) instead of playing games at different komis. But in this case we should not see those 2 exceptions. Or else I'm wrong somewhere on my assumption. You are wrong two ways. First, the exceptions are normal: they are noise. The probabilities of winning for 3.5 and 4.5 are the same. Also, you cannot deduce winning percentage from mean score. For instance +1 +1 -1 and +3 -1 -1 have the same mean score. Rémi ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Black/White winning rates with random playout?
You can't just look at the mean. If you take a histogram and look at the distribution of scores, you'll see a Gaussian-like bump in the middle, but also huge tails where only one color was left. You can calculate the histogram once and then use it to derive the win rate for different Komis. (That would save time.) But you can't use the mean for that purpose. - Dave Hillis -Original Message- From: Nuno Milheiro nunomilhe...@gmail.com To: computer-go computer-go@computer-go.org Sent: Thu, 8 Jan 2009 2:14 pm Subject: Re: [computer-go] Black/White winning rates with random playout? Given random play, komi value does not change play, so we could see what is the mean score (no komi) instead of playing games at different komis. But in this case we should not see those 2 exceptions. Or else I'm wrong somewhere on my assumption. 2009/1/8 Rémi Coulom remi.cou...@univ-lille3.fr Isaac Deutsch wrote: I ran some tests with 500k games each and came to this result: with komi 0.5, white has 47.5 winn. perc. with komi 1.5, white has 50.7 winn. perc. with komi 2.5, white has 50.9 winn. perc. with komi 3.5, white has 54.0 winn. perc. with komi 4.5, white has 53.8 winn. perc. -- ? with komi 5.5, white has 57.3 winn. perc. with komi 6.5, white has 57.1 winn. perc. - ? with komi 7.5, white has 60.3 winn. perc. with komi 8.5, white has 60.3 winn. perc. with komi 9.5, white has 63.1 winn. perc. with komi 10.5, white has 63.1 winn. perc. with komi 1 1.5, white has 65.9 winn. perc. The percentage is mostly according to komi, with 2 exceptions. If your playouts don't score seki, then komi = 2n + 1.5 and komi = 2n + 2.5 are equivalent. So you should get the same winn. perc. for 1.5 and 2.5. Ditto for 3.5 and 4.5, etc. That's because White Score + Black Score = 81, so Black Score - WhiteScore = 81 - 2 * WhiteScore is always odd. Rémi ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ -- Nuno Milheiro -BEGIN GEEK CODE BLOCK- Version: 3.12 GCM d+! s: a C ULAS$ P+ L+++$ E+++ W+ N o? K- w--$ O? M- V-- PS+++ PE- Y+ PGP++ t 5? X+ !R tv- b+ DI++ D+ G+ e+++ h r+++ y* --END GEEK CODE BLOCK-- ___ omputer-go mailing list omputer...@computer-go.org ttp://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] Black/White winning rates with random playout?
On Thu, 2009-01-08 at 20:29 +0100, Olivier Teytaud wrote: I don't know the answer, but it's not too surprising - with random play the komi should be something like 2 or 3, so white with 7.5 komi has a pretty good advantage. This advantage disappears (or almost disappears) if the games are well played, but in your case they are not. I think that the advantage increases when the games are well played, and I believe some people with other programs have the same results. In particular with good openings, white becomes very strong. Also, high level human players also have this feeling (I think). The advantage I am referring to is not real - I'm saying that 2 or 3 komi is even with random play - so if you give white more than this, he has an advantage. The advantage you are referring to is based on the assumption that 7.5 is too high and thus white (using 7.5 komi) probably has a won game technically.If white has a won game, then the stronger you are as white, the more games you will win, exactly as you say. It's also true that if, for instance, 5.5 is too low, then black has a technically won game and with stronger play black should do better and better (with strong enough play, black would win every game if 6.5 is really too low.) I know that the good bots on CGOS at 9x9 and 7.5 komi score between 50% and 51% as white.That does imply that white has a better game at 7.5 komi. Having a better game can be in a practical or real sense. In the real sense, you have a better game if it is a win with very best play, but in a practical sense it's possible that the game is won, but it's very difficult to win and may even be easier for the other side to win. Of course this depends on your skill level. With random play and 5.5 komi, I believe it is a real win for black, but for random bots the position gives white better practical chances. This is almost a paradox but it seems to be true.So if you are a random bot playing another random bot, you want to start the game with the losing side if the komi is 5.5. - Don 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/
