Re: [Computer-go] CGOS 19x19 starting trating
Yes. -Original Message- From: computer-go-boun...@dvandva.org [mailto:computer-go- boun...@dvandva.org] On Behalf Of Michael Williams Sent: Wednesday, October 13, 2010 7:40 PM To: computer-go@dvandva.org Subject: Re: [Computer-go] CGOS 19x19 starting trating Learns in what way? Won't repeat the exact same loss twice? On Wed, Oct 13, 2010 at 9:41 PM, David Fotland fotl...@smart-games.com wrote: Many Faces learns. I think most everyone follows the policy. As a developer, I want each version of my program to have a separate name so I can keep track of which results go with which version. David -Original Message- From: computer-go-boun...@dvandva.org [mailto:computer-go- boun...@dvandva.org] On Behalf Of Michael Williams Sent: Wednesday, October 13, 2010 2:48 PM To: computer-go@dvandva.org Subject: Re: [Computer-go] CGOS 19x19 starting trating It takes the burden of updating the program login name away from the programmer, which should result in the policy being followed more closely. It doesn't take away flexablility as long as the client code is changed in such away that is gracefully degrades to the current functionallity when the new field (e.g. BaseLoginName) is not present. I don't think there are any learning programs, not that that matters is you gracefully degrade to current functionallity. On Wed, Oct 13, 2010 at 8:39 AM, Don Dailey dailey@gmail.com wrote: On Wed, Oct 13, 2010 at 8:20 AM, Michael Williams michaelwilliam...@gmail.com wrote: Good points, but the suggestion wasn't meant as a method of enforcement on the policy, but instead a feature of convenience to make it easier to follow the policy. If the ratings converge fast, there is little reason to bypass the policy. But it doesn't accomplish much and takes away a little flexibility. A learning program for instance would not be solved by this for instance so it is not really particularly effective. Don On Wed, Oct 13, 2010 at 8:08 AM, Don Dailey dailey@gmail.com wrote: On Wed, Oct 13, 2010 at 7:57 AM, Michael Williams michaelwilliam...@gmail.com wrote: Yeah, but don't design the system for a case that isn't being done. You could even make this automatic in the CGOS client by specifying a base name in the config file and having the client append a hash of the executable. It's funny you say that, I once considered doing exactly that. I decided not to however as I think it's messy. it would be trivial to hack the script to send the same checksum each time and it's not always going to work anyway - if your binary is loaded via a shell script for instance you would be checksumming the script, not the actual binary. If your program is called with different command line arguments (which some of mine did) you essentially change the program without changing the binary. I decided it was too ambiguous and difficult to do right. Since it is easily circumvented it seems like the best way is an advisory system anyway. Don On Wed, Oct 13, 2010 at 7:52 AM, Don Dailey dailey@gmail.com wrote: On Wed, Oct 13, 2010 at 7:45 AM, Michael Williams michaelwilliam...@gmail.com wrote: Don, as for the program changing strength, I thought the recommended system was to create a new CGOS name when you change something in your program? I know it can't be enforeced, but just reminding people to try to stick to that policy is probably enough. Yes, that is how it should be done. The only issue here is if someone creates a program designed to learn as it plays. If we do it bayes-like then a workaround would be to periodically rename the program. On Tue, Oct 12, 2010 at 10:53 PM, Hideki Kato hideki_ka...@ybb.ne.jp wrote: http://cgos.boardspace.net/19x19/bayes.html Hideki terry mcintyre: 496924.86005...@web39803.mail.mud.yahoo.com: Is there a URL for the 19x19 bayeselo table? Terry McIntyre terrymcint...@yahoo.com Unix/Linux Systems Administration Taking time to do it right saves having to do it twice. inline file ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go -- Hideki Kato mailto:hideki_ka...@ybb.ne.jp ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org
[Computer-go] ManyFaces learning
On Wed, Oct 13, 2010 at 11:47:01PM -0700, David Fotland wrote: Many Faces learns. I think most everyone follows the policy. As a developer, I want each version of my program to have a separate name so Learns in what way? Won't repeat the exact same loss twice? Yes. Intriguing! Do you think it actually improves its strength, or is that just an experiment? How does it know where the losing variation starts? Is it updating some pattern weights, or updating some persistent game tree? Thanks, -- Petr Pasky Baudis The true meaning of life is to plant a tree under whose shade you will never sit. ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
[Computer-go] ManyFaces learning
Intriguing! Do you think it actually improves its strength, or is that just an experiment? How does it know where the losing variation starts? Is it updating some pattern weights, or updating some persistent game tree? I guess it is just an opening book that keeps win/loss info. When a branch has losses and 0 wins, you just censor that move. No need to know where the blunder was. Maybe is something more interesting. I am curious too. Jacques. ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] ManyFaces learning
That wouldn't be a very good system. Good moves would get chopped whenever you play a stronger opponent. On Thu, Oct 14, 2010 at 7:34 AM, Jacques Basaldúa jacq...@dybot.com wrote: Intriguing! Do you think it actually improves its strength, or is that just an experiment? How does it know where the losing variation starts? Is it updating some pattern weights, or updating some persistent game tree? I guess it is just an opening book that keeps win/loss info. When a branch has losses and 0 wins, you just censor that move. No need to know where the blunder was. Maybe is something more interesting. I am curious too. Jacques. ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] God Go Rank
An all-knowing opponent wouldn't play catch up in a handicap game. Reducing the the disadvantage is a heuristic for mere mortals. Instead, black would be forced into the most complicated defense possible. This means that black could actually win higher than the initial margin of roughly 3.5 * 14 in a 4 stone game. But only if he could counter perfection with perfection, which he can't. So error frequencies and average error costs are a little besides the point. Because to have a chance at winning, b would have to knowingly play suboptimal moves, trading a part of the advantage in return for reduced complexity. Stefan (Was 19x19 gets interesting.) Mark Boon commented that he believes that top pros make a few dozen errors per game. I concur, and I think that we can construct a model of Go competitions that supports that belief. with a 6.5 komi, the game is probably a forced win for Black. (It could be White, or even a draw by repetition, but that doesn't affect what follows.) So a perfect player would win 100% as Black, and also win as White unless his opponent plays perfectly. When top-ranked pros play head-to-head (e.g., in title matches) what we observe is closer to 53% wins for Black. Imagine an opponent that made an error on 1% of his turns. His probability of playing a complete game (say 130 moves) without an error is around 27%. It should be clear that such a player would win a lot more than 53% as Black. Maybe it isn't so clear. We are counting errors in point differential, so Black would win if he plays perfectly or if White makes errors that balance. If White had the same error rate as Black, then Black would win roughly 63%. So top pro error rates must exceed 1%. If the error rate were 2% then Black would still win much too often. I believe that top pros must have error rates around 9% in order to have a 53% win rate for Black. That implies a dozen errors per game per player. I am not sure what this means for dan ranking. But if 1 stone = 6 points, and 1 error = 2 points, then perfect players are about 4 stones better than top human pros. Here, top doesn't mean 9p; top means title holders. Brian ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] ManyFaces learning
If you censor at the root, you only have 361 losses that you can endure before you are out of ideas. I guess you implicitly meant censoring at move 3 or something. On Thu, Oct 14, 2010 at 7:54 AM, Jacques Basaldúa jacq...@dybot.com wrote: That wouldn't be a very good system. Good moves would get chopped whenever you play a stronger opponent. I am always thinking in 19x19. You may be right on smaller board sizes. In 19x19 games become unique really soon. The only drawback of what I suggest is that you may label a playable move as not to play. But you play another playable move. The empty 19x19 board is full of playable moves. It is only a security measure to avoid someone replaying the same game again and again against you. Censoring a good move is not a problem when there are so many. Probably in 9x9 you have to use a serious book and censoring good moves would be a problem. Jacques. ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
[Computer-go] ManyFaces learning
If you censor at the root, you only have 361 losses that you can endure before you are out of ideas. I guess you implicitly meant censoring at move 3 or something. Of course not. You start with a standard book. I my case, learned from 50K games with about 25K nodes. You are just adding learned knowledge to it. (I don't do it, btw. Just guessing what David is doing from what he posted.) Jacques. ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] ManyFaces learning
You said: When a branch has losses and 0 wins, you just censor that move. Apparently you meant leaf and not branch. On Thu, Oct 14, 2010 at 8:14 AM, Jacques Basaldúa jacq...@dybot.com wrote: If you censor at the root, you only have 361 losses that you can endure before you are out of ideas. I guess you implicitly meant censoring at move 3 or something. Of course not. You start with a standard book. I my case, learned from 50K games with about 25K nodes. You are just adding learned knowledge to it. (I don't do it, btw. Just guessing what David is doing from what he posted.) Jacques. ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
[Computer-go] ManyFaces learning
You said: When a branch has losses and 0 wins, you just censor that move. Apparently you meant leaf and not branch. Well I did not precise how you grow the tree. Obviously, the initial offline learned tree has all nodes with more than one visit as the number of visits is the criterion to build it. Then, you may stop growing the tree when you reach a branch that is unique. At that moment the last node is a leaf. So you may call it what you want. I wasn't that precise, but it makes sense. I may have explained it bad, but it can be done. Jacques. ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] God Go Rank
On Thu, Oct 14, 2010 at 01:49:49PM +0200, Stefan Kaitschick wrote: Instead, black would be forced into the most complicated defense possible. How would that be defined? One possible definition occuring to me is to always pick the variation with the minimal sum of opponent replies that lead to a win in the appropriate game subtree. But that will just give huge preference to fights with the biggest amount of forcing moves. Petr Pasky Baudis ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] ManyFaces learning
Ideally, a program could begin at the end of a lost game, and work back to determine at what move it lost the game. This is usually not the last move, but some distance back; it may even be in the fuseki stage. If this offline analysis is complete, it could be used to confidently censor bad moves. Never make that mistake again. The next stage in learning would be to analyze why the bad decision was made, and to search for a way to improve the decision-making process. Did the program not discover a clever tesuji by the opponent? Did it mis-read the life and death status of a group? Fail to respond appropriately in a capturing race? Automating this sort of learning would be an interesting task. When a bad move is censored, one could also strike variations - rotations of the board, for example. If the program has good analytical skills, it could create advice such as don't play this if the ladder doesn't work. ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
[Computer-go] TAAI2010 computer game tournament(11/18 and 11/19, Taiwan)
TAAI 2010 computer game tournament will be held at November 18th and 19th at National Chiao Tung University in Hsin Chu, Taiwan, with TAAI2010. For more information, please visit http://ai.csie.ndhu.edu.tw:9898/eng/ and http://taai2010.nctu.edu.tw/ Go 9x9, 13x13 and 19x19 will be played. Erica, Mogo and Zen attended TAAI2009 last year. GoIntelligent and Mogo have registered this year. The rules are similar to computer Olympiad, but will be played on KGS. Best regards, Shi-Jim Yen Associate Professor of CSIE, in NDHU, Houlian, Taiwan. ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
[Computer-go] CGOS standard engine proposal
It makes sense to have more than one standard program. Preferably covering a wide range of skills. There is no need to fix the ratings of standard programs, provided that they play a large number of games. The problem is to find the resources to host a wide range of standard players. I experienced a period on CGOS where 3 programs were stronger than Pebbles by 300 rating points, and the standard 9x9 programs were weaker by 300 rating points, and no opponents were in Pebbles range. I resolved that by running two copies of Pebbles, and by creating a fixed version of Fuego running on a spare computer. Now there was always a good matchup to be found. But that solution only works when Pebbles is playing. I feel like a collaborative solution should work. Here is a proposal: 1) We adopt a naming convention for standard players. 2) We create a set of standard players, scaled over a range from far below the anchor to far above. 3) CGOS blocks creating names that follow the convention except by the server admin. 4) We publish the commands that execute the standard players. 5) Anyone can create an instance of a standard version (provided that no other user is currently running an instance). 6) Now you can run instances of standards that scale to the rating range you care about. For example, suppose Fuego will be a standard. Our naming convention could be Std-Fuego-0.4-2000, to indicate that the player is a standard player, running Fuego, version 0.4, fixed to 2000 trials per turn. It makes sense to have a diversity of programs, so that we don't get in the rut of tuning to a specific program's weaknesses. Open source engines like Fuego, Pachi, and Orego would make excellent choices. We would need to archive binary versions that the standards depend upon. Thoughts? Brian ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
[Computer-go] Pebbles opening learning
The Mogo team (among others) published extensive descriptions of how to write an opening library that learns. I recommend following their idea, rather than mine. :-) But if you don't have the time to do things properly, and you don't want to lose the exact same game over and over, then you might use Pebbles simple adaptive technique. Pebbles repeats moves that won the last time they were played. Specifically: 1) after every game, mark all positions in which the winner moved as winning-last-time. 2) In every game look up all successors of the current position. 3) Choose randomly among successors that were winning-last-time. 4) If no winning-last-time then search. This simple system is surprisingly robust. In particular, a) your program learns from opponents that defeat it. b) your program repeats moves that previously won. c) The selected moves have a strength that is lower-bounded by your program's native search strength, because i) A move either came from your own search, or ii) from a previous game in which an opponent defeated you. The weakness is in step 4: search. It is possible to have a winning position, but your search can't find the play. It is possible to upgrade step 4 so that the process is asymptotically optimal. That is, the system converges on maximal play. Pebbles hasn't actually implemented that feature, but I will have to implement that eventually in order to climb the ladder. Brian ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] God Go Rank
Minimal number of opponent replies is the right idea. But not just the next move. The whole structure of the following game tree is important. There is also the consideration that successful b moves should not be easy to find. But maybe this would be playing with the devil. :-) Stefan On Thu, Oct 14, 2010 at 01:49:49PM +0200, Stefan Kaitschick wrote: Instead, black would be forced into the most complicated defense possible. How would that be defined? One possible definition occuring to me is to always pick the variation with the minimal sum of opponent replies that lead to a win in the appropriate game subtree. But that will just give huge preference to fights with the biggest amount of forcing moves. Petr Pasky Baudis ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] God Go Rank
Sorry, I didn't read your post properly. You already considered the full game tree. Probably, minimal product is slightly better than minimal sum. I'm not sure what the statistical properties of a wild fullboard fight would be. But such a superior player should have no problem discerning that. Stefan On Thu, Oct 14, 2010 at 01:49:49PM +0200, Stefan Kaitschick wrote: Instead, black would be forced into the most complicated defense possible. How would that be defined? One possible definition occuring to me is to always pick the variation with the minimal sum of opponent replies that lead to a win in the appropriate game subtree. But that will just give huge preference to fights with the biggest amount of forcing moves. Petr Pasky Baudis ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] TAAI2010 computer game tournament(11/18 and 11/19, Taiwan)
Can we play remotely, by connecting to KGS, or does the program have to run on a computer in Taiwan at the contest site? Thanks, David From: computer-go-boun...@dvandva.org [mailto:computer-go-boun...@dvandva.org] On Behalf Of Shi-Jim Yen Sent: Thursday, October 14, 2010 6:34 AM To: computer-go@dvandva.org Subject: [Computer-go] TAAI2010 computer game tournament(11/18 and 11/19, Taiwan) TAAI 2010 computer game tournament will be held at November 18th and 19th at National Chiao Tung University in Hsin Chu, Taiwan, with TAAI2010. For more information, please visit http://ai.csie.ndhu.edu.tw:9898/eng/ and http://taai2010.nctu.edu.tw/ Go 9x9, 13x13 and 19x19 will be played. Erica, Mogo and Zen attended TAAI2009 last year. GoIntelligent and Mogo have registered this year. The rules are similar to computer Olympiad, but will be played on KGS. Best regards, Shi-Jim Yen Associate Professor of CSIE, in NDHU, Houlian, Taiwan. ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] CGOS standard engine proposal
If there are multiple anchors, then what should be fixed is the average rating of them assuming they all stay active. But there still needs to be something fixed or else the rating pool will be subject to random drift. There is no limit to the number of possible ideas to deal with this, but I'm not currently in a position to work on CGOS, so I prefer something that is simple and compatible with what we already have.The simplest idea (if we really think we need a change) is to upgrade our anchor player. The rating issue could be solved in a simple way - only publish the bayeselo ratings of the players. Internally I can configure the server to use a higher K factor so that it more quickly adapts but the rating we publish is Bayeselo. On the server Bayeselo is updated only every 6 hours, so you would not get rating updates after every game.Another minor downside is that a bot's rating can change even when it's not playing games. To me this is not a bad thing because newer ratings will more accurately reflect the strength of the programs. Don On Thu, Oct 14, 2010 at 11:03 AM, Brian Sheppard sheppar...@aol.com wrote: It makes sense to have more than one standard program. Preferably covering a wide range of skills. There is no need to fix the ratings of standard programs, provided that they play a large number of games. The problem is to find the resources to host a wide range of standard players. I experienced a period on CGOS where 3 programs were stronger than Pebbles by 300 rating points, and the standard 9x9 programs were weaker by 300 rating points, and no opponents were in Pebbles range. I resolved that by running two copies of Pebbles, and by creating a fixed version of Fuego running on a spare computer. Now there was always a good matchup to be found. But that solution only works when Pebbles is playing. I feel like a collaborative solution should work. Here is a proposal: 1) We adopt a naming convention for standard players. 2) We create a set of standard players, scaled over a range from far below the anchor to far above. 3) CGOS blocks creating names that follow the convention except by the server admin. 4) We publish the commands that execute the standard players. 5) Anyone can create an instance of a standard version (provided that no other user is currently running an instance). 6) Now you can run instances of standards that scale to the rating range you care about. For example, suppose Fuego will be a standard. Our naming convention could be Std-Fuego-0.4-2000, to indicate that the player is a standard player, running Fuego, version 0.4, fixed to 2000 trials per turn. It makes sense to have a diversity of programs, so that we don't get in the rut of tuning to a specific program's weaknesses. Open source engines like Fuego, Pachi, and Orego would make excellent choices. We would need to archive binary versions that the standards depend upon. Thoughts? Brian ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] ManyFaces learning
On 9x9, playing against people, the opening learning is very important so that they dont find a winning sequence then just repeat it. This has been in Igowin since it first shipped, in 1998. I dont think it helps the strength much in 19x19, but it makes the games look prettier. I use a persistent position hash table that stores every position visited twice (or more) by games that were deliberately added (File, Add Game(s) to database), and every position in every game played locally. The program ships with a 19x19 database built from about 40K games. The next update will also include 9x9 games from strong cgos programs. Move choices are biased based on the win/loss ratio, number of visits, strength of strongest player moving into that position, depth in the game, and probably more that I have long since forgotten. There is no hard pruning or immediate move choices from this data, just biasing the search. David -Original Message- From: computer-go-boun...@dvandva.org [mailto:computer-go- boun...@dvandva.org] On Behalf Of Petr Baudis Sent: Thursday, October 14, 2010 3:02 AM To: computer-go@dvandva.org Subject: [Computer-go] ManyFaces learning On Wed, Oct 13, 2010 at 11:47:01PM -0700, David Fotland wrote: Many Faces learns. I think most everyone follows the policy. As a developer, I want each version of my program to have a separate name so Learns in what way? Won't repeat the exact same loss twice? Yes. Intriguing! Do you think it actually improves its strength, or is that just an experiment? How does it know where the losing variation starts? Is it updating some pattern weights, or updating some persistent game tree? Thanks, -- Petr Pasky Baudis The true meaning of life is to plant a tree under whose shade you will never sit. ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] CGOS standard engine proposal
If there are multiple anchors I didn't propose multiple anchors, just multiple standard players. Just to clarify, I proposed keeping the existing anchors (FatMan on 9 and GnuGo on 13 and 19). Then define a set of standard players that *anyone* can run. This way, if I want an opponent around 2200 then I can start an appropriately chosen standard. Because of the collaborative efforts of all developers who have had that need, such a player would have thousands of games and a very accurate rating. This system relieves you of the burden of running multiple standards, while allowing the community to provide the computing power that supports its needs over time. My proposal does not require a code change. Just a recognition that we can establish specific points on a rating scale by convention, using collaborative computing. Brian ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] CGOS standard engine proposal
Brian Sheppard wrote: I feel like a collaborative solution should work. Here is a proposal: 1) We adopt a naming convention for standard players. 2) We create a set of standard players, scaled over a range from far below the anchor to far above. 3) CGOS blocks creating names that follow the convention except by the server admin. 4) We publish the commands that execute the standard players. 5) Anyone can create an instance of a standard version (provided that no other user is currently running an instance). I think that's a very practical idea. I'd be happy to run a standard player or two from time to time. -M- ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] CGOS starting rating
The rating issue could be solved in a simple way - only publish the bayeselo ratings of the players. (Don posted this under a different thread, but I think it belongs here.) The proposal to use bayeselo to rate players does not work for players that learn, as Pebbles does. Pebbles bayeselo trails its incremental rating by hundreds of points. If you publish Bayeselo, then people will wonder why Pebbles wins so much. Brian ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] CGOS standard engine proposal
I also like this suggestion. Funnily enough I was thinking that it would be better to have more than one standard player earlier today! I like Don's list of requirements for the standard algorithms. I guess this cuts the list of possibles down a bit? If you want a holding place for suggesting standard players, feel free to add a page to the wiki Matthew Woodcraft wrote: Brian Sheppard wrote: I feel like a collaborative solution should work. Here is a proposal: 1) We adopt a naming convention for standard players. 2) We create a set of standard players, scaled over a range from far below the anchor to far above. 3) CGOS blocks creating names that follow the convention except by the server admin. 4) We publish the commands that execute the standard players. 5) Anyone can create an instance of a standard version (provided that no other user is currently running an instance). I think that's a very practical idea. I'd be happy to run a standard player or two from time to time. -M- ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] ManyFaces learning
I don’t think it helps the strength much in 19x19, but it makes the games look prettier. So regarding learning while playing on CGOS: The program changes, but the amount that the true ELO is going to move during it's CGOS lifetime is completely negligable. On Thu, Oct 14, 2010 at 1:09 PM, David Fotland fotl...@smart-games.com wrote: On 9x9, playing against people, the opening learning is very important so that they don’t find a winning sequence then just repeat it. This has been in Igowin since it first shipped, in 1998. I don’t think it helps the strength much in 19x19, but it makes the games look prettier. I use a persistent position hash table that stores every position visited twice (or more) by games that were deliberately added (File, Add Game(s) to database), and every position in every game played locally. The program ships with a 19x19 database built from about 40K games. The next update will also include 9x9 games from strong cgos programs. Move choices are biased based on the win/loss ratio, number of visits, strength of strongest player moving into that position, depth in the game, and probably more that I have long since forgotten. There is no hard pruning or immediate move choices from this data, just biasing the search. David -Original Message- From: computer-go-boun...@dvandva.org [mailto:computer-go- boun...@dvandva.org] On Behalf Of Petr Baudis Sent: Thursday, October 14, 2010 3:02 AM To: computer-go@dvandva.org Subject: [Computer-go] ManyFaces learning On Wed, Oct 13, 2010 at 11:47:01PM -0700, David Fotland wrote: Many Faces learns. I think most everyone follows the policy. As a developer, I want each version of my program to have a separate name so Learns in what way? Won't repeat the exact same loss twice? Yes. Intriguing! Do you think it actually improves its strength, or is that just an experiment? How does it know where the losing variation starts? Is it updating some pattern weights, or updating some persistent game tree? Thanks, -- Petr Pasky Baudis The true meaning of life is to plant a tree under whose shade you will never sit. ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] ManyFaces learning
On 19x19 yes, but on 9x9, it probably can get noticeably stronger, especially if other strong programs have openings they like and MF learns a refutation. David -Original Message- From: computer-go-boun...@dvandva.org [mailto:computer-go- boun...@dvandva.org] On Behalf Of Michael Williams Sent: Thursday, October 14, 2010 3:03 PM To: computer-go@dvandva.org Subject: Re: [Computer-go] ManyFaces learning I dont think it helps the strength much in 19x19, but it makes the games look prettier. So regarding learning while playing on CGOS: The program changes, but the amount that the true ELO is going to move during it's CGOS lifetime is completely negligable. On Thu, Oct 14, 2010 at 1:09 PM, David Fotland fotl...@smart-games.com wrote: On 9x9, playing against people, the opening learning is very important so that they dont find a winning sequence then just repeat it. This has been in Igowin since it first shipped, in 1998. I dont think it helps the strength much in 19x19, but it makes the games look prettier. I use a persistent position hash table that stores every position visited twice (or more) by games that were deliberately added (File, Add Game(s) to database), and every position in every game played locally. The program ships with a 19x19 database built from about 40K games. The next update will also include 9x9 games from strong cgos programs. Move choices are biased based on the win/loss ratio, number of visits, strength of strongest player moving into that position, depth in the game, and probably more that I have long since forgotten. There is no hard pruning or immediate move choices from this data, just biasing the search. David -Original Message- From: computer-go-boun...@dvandva.org [mailto:computer-go- boun...@dvandva.org] On Behalf Of Petr Baudis Sent: Thursday, October 14, 2010 3:02 AM To: computer-go@dvandva.org Subject: [Computer-go] ManyFaces learning On Wed, Oct 13, 2010 at 11:47:01PM -0700, David Fotland wrote: Many Faces learns. I think most everyone follows the policy. As a developer, I want each version of my program to have a separate name so Learns in what way? Won't repeat the exact same loss twice? Yes. Intriguing! Do you think it actually improves its strength, or is that just an experiment? How does it know where the losing variation starts? Is it updating some pattern weights, or updating some persistent game tree? Thanks, -- Petr Pasky Baudis The true meaning of life is to plant a tree under whose shade you will never sit. ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] CGOS standard engine proposal
Robert, could you use plain text mail please? I run several such standards as well as the anchors, gnugo-3.8-l10F and gnugo-3.8-mc in addition to FatMan-1 on 9x9 and GNU-Go-3.8-L12 on 19x19, for long time. Also I had run Zengg-2x4c again to pull up strong new programs. The idea that volunteers run many standards is good and there are no reason to against but I'd like to suggest don't expect much. That will accelerate the convergence, most likely, ten times or around, depending on the interval they will be placed (assuming 200 Elo) but much slower than copying bayeselo which will be updated every six hours. Possible drawbacks are, not ecological :) and since the standard programs have to be public and limited to a few ones such as Fuego and MoGo, the ratings will be biased in some sense. Hideki Robert Finking: 4cb76ebf.2080...@o2.co.uk: I also like this suggestion. Funnily enough I was thinking that it would be better to have more than one standard player earlier today! I like Don's list of requirements for the standard algorithms. I guess this cuts the list of possibles down a bit? If you want a holding place for suggesting standard players, feel free to add a page to the wiki Matthew Woodcraft wrote: Brian Sheppard wrote: I feel like a collaborative solution should work. Here is a proposal: 1) We adopt a naming convention for standard players. 2) We create a set of standard players, scaled over a range from far below the anchor to far above. 3) CGOS blocks creating names that follow the convention except by the server admin. 4) We publish the commands that execute the standard players. 5) Anyone can create an instance of a standard version (provided that no other user is currently running an instance). I think that's a very practical idea. I'd be happy to run a standard player or two from time to time. -M- -- Hideki Kato mailto:hideki_ka...@ybb.ne.jp ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] CGOS standard engine proposal
That will accelerate the convergence, most likely, ten times or around, depending on the interval they will be placed (assuming 200 Elo) but much slower than copying bayeselo which will be updated every six hours. Convergence is an independent issue. Speeding convergence is useful, but not the only reason for wanting to define standard players. First, the pool can be thin. You will often see fewer than 8 programs playing. It would help to start another engine. Why not start a standard? Second, the pool can be thin near your program. Pebbles plays in a region that is often underrepresented. Third, a standard offers some reproducibility. If you lose against a standard, and want to reproduce, then you can. Fourth, standards allow you to test offline against a player of known rating. Fifth, a family of standard players that differ by speed establishes rating curves. You can do the bayeselo, too, if you want. Brian ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] CGOS standard engine proposal
Dear Brian, Hideki Kato wrote in 4cb79ef8.1360%hideki_ka...@ybb.ne.jp: The idea that volunteers run many standards is good and there are no reason to against but I'd like to suggest don't expect much. Hideki Brian Sheppard: dee6fa7b72d947b5a2b6474bf842f...@inspirone1705: That will accelerate the convergence, most likely, ten times or around, depending on the interval they will be placed (assuming 200 Elo) but much slower than copying bayeselo which will be updated every six hours. Convergence is an independent issue. Speeding convergence is useful, but not the only reason for wanting to define standard players. First, the pool can be thin. You will often see fewer than 8 programs playing. It would help to start another engine. Why not start a standard? Second, the pool can be thin near your program. Pebbles plays in a region that is often underrepresented. Third, a standard offers some reproducibility. If you lose against a standard, and want to reproduce, then you can. Fourth, standards allow you to test offline against a player of known rating. Fifth, a family of standard players that differ by speed establishes rating curves. You can do the bayeselo, too, if you want. Brian ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go -- Hideki Kato mailto:hideki_ka...@ybb.ne.jp ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] CGOS standard engine proposal
I just put up mfgo12-636 (originally from March 2010), which was the highest rated Many Faces until recently. I'll keep it up as I can to give another highly rated reference. It ended up rated quite a bit higher than programs that are stronger now, but have lower ratings. I think it might be a good idea to relax the final K factor a little so programs with hundreds of games can still move their ratings a little faster. David -Original Message- From: computer-go-boun...@dvandva.org [mailto:computer-go- boun...@dvandva.org] On Behalf Of Hideki Kato Sent: Thursday, October 14, 2010 5:23 PM To: computer-go@dvandva.org Subject: Re: [Computer-go] CGOS standard engine proposal Robert, could you use plain text mail please? I run several such standards as well as the anchors, gnugo-3.8-l10F and gnugo-3.8-mc in addition to FatMan-1 on 9x9 and GNU-Go-3.8-L12 on 19x19, for long time. Also I had run Zengg-2x4c again to pull up strong new programs. The idea that volunteers run many standards is good and there are no reason to against but I'd like to suggest don't expect much. That will accelerate the convergence, most likely, ten times or around, depending on the interval they will be placed (assuming 200 Elo) but much slower than copying bayeselo which will be updated every six hours. Possible drawbacks are, not ecological :) and since the standard programs have to be public and limited to a few ones such as Fuego and MoGo, the ratings will be biased in some sense. Hideki Robert Finking: 4cb76ebf.2080...@o2.co.uk: I also like this suggestion. Funnily enough I was thinking that it would be better to have more than one standard player earlier today! I like Don's list of requirements for the standard algorithms. I guess this cuts the list of possibles down a bit? If you want a holding place for suggesting standard players, feel free to add a page to the wiki Matthew Woodcraft wrote: Brian Sheppard wrote: I feel like a collaborative solution should work. Here is a proposal: 1) We adopt a naming convention for standard players. 2) We create a set of standard players, scaled over a range from far below the anchor to far above. 3) CGOS blocks creating names that follow the convention except by the server admin. 4) We publish the commands that execute the standard players. 5) Anyone can create an instance of a standard version (provided that no other user is currently running an instance). I think that's a very practical idea. I'd be happy to run a standard player or two from time to time. -M- -- Hideki Kato mailto:hideki_ka...@ybb.ne.jp ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Pebbles opening learning
I presume this is for small boards? Hard to see in 19x19 that doing such statistics would be meaningful. Petri 2010/10/14 Brian Sheppard sheppar...@aol.com The Mogo team (among others) published extensive descriptions of how to write an opening library that learns. I recommend following their idea, rather than mine. :-) But if you don't have the time to do things properly, and you don't want to lose the exact same game over and over, then you might use Pebbles simple adaptive technique. Pebbles repeats moves that won the last time they were played. Specifically: 1) after every game, mark all positions in which the winner moved as winning-last-time. 2) In every game look up all successors of the current position. 3) Choose randomly among successors that were winning-last-time. 4) If no winning-last-time then search. This simple system is surprisingly robust. In particular, a) your program learns from opponents that defeat it. b) your program repeats moves that previously won. c) The selected moves have a strength that is lower-bounded by your program's native search strength, because i) A move either came from your own search, or ii) from a previous game in which an opponent defeated you. The weakness is in step 4: search. It is possible to have a winning position, but your search can't find the play. It is possible to upgrade step 4 so that the process is asymptotically optimal. That is, the system converges on maximal play. Pebbles hasn't actually implemented that feature, but I will have to implement that eventually in order to climb the ladder. Brian ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
