Re: [computer-go] Elo and handicap matching
Alain Baeckeroot wrote: More subbtle attempts are summarized at: http://senseis.xmp.net/?HandicapForSmallerBoardSizes The previously suggested komi there were terrible. I have added and justified my suggestions: 8.5 for 13x13, 6.5 for 9x9. -- robert jasiek ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] erm...
Don Dailey wrote: > you can still judge the quality of your opponent by looking at his 19x19 KGS ranking. Rather by looking at his real world ranking. A human real world rank may be off by 1 while a human KGS rank may be off by 6 ranks. -- robert jasiek ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
Hi Christoph, I have been thinking about making a version of Goanna (~2250 on CGOS) public, once it plays in a human friendly way. At the moment, it is nearly unusable for fun human vs computer matches because of a lack of opening book (slow first few moves), and ridiculous endgame play. Considering how much time I have been putting into this project lately, it is not going to be happening till at least the end of January. Joel On Dec 5, 2007 8:13 AM, Christoph Birk <[EMAIL PROTECTED]> wrote: > On Tue, 4 Dec 2007, Gunnar Farnebäck wrote: > > terry mcintyre wrote: > >> Some of the MonteGNU code was just released on CVS. Check out Gnugo's > >> development pages. > > > > Don't expect that code to do better than 2000 on CGOS though > > (mgtest2). The remaining code used by MonteGNU is still too messy. > > That's why I asked for 'MonteGNU'. It has a well established rating > on CGOS (23670 games). > I there any other (fairly) strong CGOS 9x9 program that's available > for download AND that has an reliable rating on CGOS? > > Christoph > > ___ > 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] programs at the US Go Congress
On Tue, 2007-12-04 at 18:23 -0800, David Doshay wrote: > What would get YOU to bring your program to the Congress? Free trips back and forth on a teleporter. Or at least 3 unlikely events (out of the US Go Congress's control) to occur. It's probably be more viable for people to send their programs than to show up in person. It's probably possible to grab a few windows computers and put bots on them. My program is weak enough that I don't mind running on inferior hardware. > What would > you like to do once you are there? Honestly, the previous event really had little non-go stuff to do. I just can't play go or watch go lectures all day. For the previous one, I was able to drive 2 hours to get there. I spent only 3 hours or so at the event (as a non-player) and then turned around to go home. Another gathering of computer go developers. There was one at the last event, but the organization of it was horrible. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] erm...
Darren Cook wrote: >> 9x9 games is a bit silly. it doesn't actually capture any extra >> information about the program, since there's no such thing as >> a 9x9 rank to compare with/against, much less a dan rank. >> > > I disagree. In my studies of 9x9, over a number of years, the human > 19x19 rank generally carries over to 9x9. E.g. a 9p consistently beats a > 3p, a 6d consistently beats a 3d, a 1d consistently beats a 5k. Whether > they are playing at 19x19 or 9x9. > That's my assumption too. Although it's possible some players are better at one board size than another I don't believe the effect is likely to be major. Even if a strong player never played a 9x9 game I believe he would quickly adapt to it after 2 or 3 games. > (As an aside, my conclusion from this, which I personally think is very > important, is that the main element of human go strength is shape, > tesuji and life/death reading.) > > UCT programs are a bit different, in that (as they are "brute-forcing" > in the number of legal moves) their strength does not scale up to higher > board sizes in the same way human strength does. But no-one is claiming > otherwise which is why people are referring to the 9x9 strength and the > 19x19 strength of Mogo and CrazyStone as separate things. > It's interesting to me that 19x19 go programs start out weaker relative to humans than 9x9 go programs. But I ran experiments that indicated wonderful scaling behavior in 19x19 go. If anything they improved MORE with depth. Of course I could only test at the low end of the skill range with a weak UCT program (an older Lazarus version.) The game at 19x19 is more difficult for computers, but this is very true of humans too. It's possible that strong humans just don't specialize in 9x9 go so perhaps this is a partial explanation.I think the good UCT programs are about 2-3 kyu lower at 19x19 relative to humans. Could it be that really it's the humans weaker at 9x9 since strong humans don't really specialize in 9x9 go? - Don > But, when we say Mogo is 3d at 9x9, it is fair to say that a human > player with a 3d rank at 19x19 will enjoy a challenging game with it. > > 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/
Re: [computer-go] New engine? From a Chess programmer perspective.
>> I've played the top 9x9 programs at 9x9, and so have several >> other amateur Dan players, and I think we all agree that the top 9x9 >> programs have reached amateur Dan level. I'll add another vote for that opinion. (3-dan-ish, at 30-60s/move, on a 2.8Ghz Celeron). Robert, you can get Mogo here, and I'm sure I'm not the only one who would be interested to hear the results of your games against it: http://www.lri.fr/~gelly/MoGo_Download.htm Strength will increase the more CPU cycles it is given, so please also report back the time limits and CPU cores & speed you are playing at. Darren ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] erm...
> 9x9 games is a bit silly. it doesn't actually capture any extra > information about the program, since there's no such thing as > a 9x9 rank to compare with/against, much less a dan rank. I disagree. In my studies of 9x9, over a number of years, the human 19x19 rank generally carries over to 9x9. E.g. a 9p consistently beats a 3p, a 6d consistently beats a 3d, a 1d consistently beats a 5k. Whether they are playing at 19x19 or 9x9. (As an aside, my conclusion from this, which I personally think is very important, is that the main element of human go strength is shape, tesuji and life/death reading.) UCT programs are a bit different, in that (as they are "brute-forcing" in the number of legal moves) their strength does not scale up to higher board sizes in the same way human strength does. But no-one is claiming otherwise which is why people are referring to the 9x9 strength and the 19x19 strength of Mogo and CrazyStone as separate things. But, when we say Mogo is 3d at 9x9, it is fair to say that a human player with a 3d rank at 19x19 will enjoy a challenging game with it. Darren ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] programs at the US Go Congress
Hi, What should we be doing to get programmers to bring their bots to the Congress in Portland in 2008? The AGA is formally not that program friendly, but there can be events for bots, and hopefully events for humans against bots. I am sure that there will be tournaments that will not include programs, but if we start now I am sure that we can make it interesting for all. Chuck Robins, one of the TDs, said that he is willing to have tournaments that allow computers. The recent thread on testing the strong programs and what ratings they might have has me thinking that it might be good to have various programs set up to play against any person who approaches the computer. If we develop a method to keep track of the ranks of the opponents, by the end of the Congress each program will have an aprox rating. What would get YOU to bring your program to the Congress? What would you like to do once you are there? Cheers, David ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
You must also avoid suicide moves! I also tried playing on top of an existing stone and it didn't allow that - but any other kind of illegal move (by cgos rules) is passed through and causes a CGOS forfeit. There is a config file option, perhaps there is way to configure it to a particular set of rules? - Don Don Dailey wrote: > I just tried gtpdisplay and it worked the first time!The only > problem is that I tried to make an illegal ko move. > > On linux, I just put gtpdisplay as the name of the program and it > worked. > > It looks like it could also be used to watch your program play on CGOS, > just provide a program name as an argument to the program. > > I don't see any way to make it reject superko however - so you must take > care not to make a superko move. > > - Don > > > > > Christoph Birk wrote: > >> On Tue, 4 Dec 2007, Don Dailey wrote: >> Yes, that would work. Some humans also could play on CGOS (just for a while) to establish a conversion from CGOS-ELO to human-ranks. >>> It would be awkward at best. I could build a client to do this, but >>> the human would have to be willing to sit and play games at the moment >>> they were scheduled. >>> >>> It would be better to get a handful of players to AGREE in advance to >>> play particular programs on KGS but under very CGOS like conditions >>> including the time control. They should know in advance they are >>> part of the experiment and that they should not experiment but should >>> play full strength and not abandon the games or take-back moves, etc. >>> >> Yes, this would be better ... but will it happen? >> I quickly hacked my program to allow me to make moves >> and I am playing currently as test-3k (my AGA rating) >> (so much for 'C' beeing slow for prototyping :-) >> >> Christoph >> >> ___ >> 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] Where and How to Test the Strong Programs?
I just tried gtpdisplay and it worked the first time!The only problem is that I tried to make an illegal ko move. On linux, I just put gtpdisplay as the name of the program and it worked. It looks like it could also be used to watch your program play on CGOS, just provide a program name as an argument to the program. I don't see any way to make it reject superko however - so you must take care not to make a superko move. - Don Christoph Birk wrote: > On Tue, 4 Dec 2007, Don Dailey wrote: >>> Yes, that would work. >>> Some humans also could play on CGOS (just for a while) to establish >>> a conversion from CGOS-ELO to human-ranks. >>> >> It would be awkward at best. I could build a client to do this, but >> the human would have to be willing to sit and play games at the moment >> they were scheduled. >> >> It would be better to get a handful of players to AGREE in advance to >> play particular programs on KGS but under very CGOS like conditions >> including the time control. They should know in advance they are >> part of the experiment and that they should not experiment but should >> play full strength and not abandon the games or take-back moves, etc. > > Yes, this would be better ... but will it happen? > I quickly hacked my program to allow me to make moves > and I am playing currently as test-3k (my AGA rating) > (so much for 'C' beeing slow for prototyping :-) > > Christoph > > ___ > 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] GTP back to basics.
Thanks that's helps the most. Also appreciate the post on how to search the comp-go archives using google. Still hoping to have olympus on cgos by the end of this month :) -Josh On Dec 4, 2007 8:34 PM, Don Dailey <[EMAIL PROTECTED]> wrote: > If you download cgos3.tcl, the code made it pretty clear which commands > are used. > > Here is what I see: > >list_commands >boardsize >clear_board >play >genmove >quit > >optional are: >time_settings >time_left > > > -Don > > > > > > Joshua Shriver wrote: > > Sorry to duplicate my question, I've been digging and digging in > > previous threads trying to find the answers that were posted. > > Wish computer-go had a google search :) > > > > Anyway, what is the minimal commands required to get an engine online > > via gtp? > > > > I've been working hard and hope to have an alpha program on CGOS by > > Christmas. Know last time I asked this, one useful reply gave just 4-5 > > commands that were essential. > > > > I have the full spec, and nothing look terribly hard, but I want to > > get the minimal commands done first so I can get my program online asap. > > > > Thanks in advance! > > -Josh > > > > > > ___ > > 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] GTP back to basics.
If you download cgos3.tcl, the code made it pretty clear which commands are used. Here is what I see: list_commands boardsize clear_board play genmove quit optional are: time_settings time_left -Don Joshua Shriver wrote: > Sorry to duplicate my question, I've been digging and digging in > previous threads trying to find the answers that were posted. > Wish computer-go had a google search :) > > Anyway, what is the minimal commands required to get an engine online > via gtp? > > I've been working hard and hope to have an alpha program on CGOS by > Christmas. Know last time I asked this, one useful reply gave just 4-5 > commands that were essential. > > I have the full spec, and nothing look terribly hard, but I want to > get the minimal commands done first so I can get my program online asap. > > Thanks in advance! > -Josh > > > ___ > 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] GTP back to basics.
On Tue, 2007-12-04 at 19:45 -0500, Joshua Shriver wrote: > Wish computer-go had a google search :) Put this in Google's search box: site:http://computer-go.org/pipermail/computer-go/ foobar Of course, replace "foobar" with your search terms. -Jeff ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] xboard like for Go
GoGui is written in Java. So you should be able to use it in Linux. On Dec 4, 2007 7:36 PM, Joshua Shriver <[EMAIL PROTECTED]> wrote: > Anyone recommend a good Go GUI for Linux? Not for bot matches and suchs but > just to play gtp based engines. > For chess I use xboard and it's wondeful, would love to find a similiar tool > for Go. > > -Josh > > ___ > 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] GTP back to basics.
Sorry to duplicate my question, I've been digging and digging in previous threads trying to find the answers that were posted. Wish computer-go had a google search :) Anyway, what is the minimal commands required to get an engine online via gtp? I've been working hard and hope to have an alpha program on CGOS by Christmas. Know last time I asked this, one useful reply gave just 4-5 commands that were essential. I have the full spec, and nothing look terribly hard, but I want to get the minimal commands done first so I can get my program online asap. Thanks in advance! -Josh ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] xboard like for Go
Anyone recommend a good Go GUI for Linux? Not for bot matches and suchs but just to play gtp based engines. For chess I use xboard and it's wondeful, would love to find a similiar tool for Go. -Josh ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Elo and handicap matching
Le mardi 4 décembre 2007, Christoph Birk a écrit : > On Tue, 4 Dec 2007, Alain Baeckeroot wrote: > >> For 9x9 ELO works better. For 19x19 it's less clear cut.The > >> handicap system appears to be a good system at 19x19 and has the very > >> nice merit of allowing grossly mismatched players to compete. I > >> think the two systems can be married by adding a fixed offset per stone > >> handicap to your ELO. > >> > > > > Stats from official european go federation, on 150 000 games nearly solves > > the problem of doing ELO vs handicap matching. > > http://gemma.ujf.cas.cz/~cieply/GO/statev.html > > > > Just need to find one anchor, lets says gnugo , rated 6k on kgs in > > 2007-11... > > That's nice, but on 19x19. I was interested in calibrating > CGOS 9x9 versus some moderatly strong humans. It is possible to download mogo games on kgs and filter dan players games then make some stats. My estimate was more than 2d (kgs) for mogo 1.0 some monthes ago. My experience on 9x9 on kgs against mogo_1.0: I was 1 kyu on kgs and mogo was significantly stronger than me: i lose more than 2/3 on 10 games, and the rare victories were due to some blunder of mogo (nakade in corner, instead of seki). Alain ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Elo and handicap matching
Le mardi 4 décembre 2007, Don Dailey a écrit : > The only issue is that I don't know if GnuGo is representative of 19x19 > to 9x9 go strength. I am interested in knowing how a human 19x19 > scales down to 9x9 play. It's well known that programs scale up poorly. Ah yes, i forgot this :) My personnal rule is - on 13x13 divide handicap by 2 - on 9x9 divide by 3 It is a bit rough, but quite good. More subbtle attempts are summarized at: http://senseis.xmp.net/?HandicapForSmallerBoardSizes http://senseis.xmp.net/?AGAHandicaps and http://www.timhunt.me.uk/go/handicaps/ with and "interesting observation" in the end. Alain. > > However, this data should still be quite useful. > > - Don > > > Alain Baeckeroot wrote: > > Le mardi 4 décembre 2007, Don Dailey a écrit : > > > > > >> For 9x9 ELO works better. For 19x19 it's less clear cut.The > >> handicap system appears to be a good system at 19x19 and has the very > >> nice merit of allowing grossly mismatched players to compete. I > >> think the two systems can be married by adding a fixed offset per stone > >> handicap to your ELO. > >> > >> > > > > Stats from official european go federation, on 150 000 games nearly solves > > the problem of doing ELO vs handicap matching. > > http://gemma.ujf.cas.cz/~cieply/GO/statev.html > > > > Just need to find one anchor, lets says gnugo , rated 6k on kgs in > > 2007-11... > > > > Alain > > > > PS i posted this link some times on the list, but nobody seems to > > consider it is useful (except Sylvain Gelly ;-) > > I would be glad if someone could explain to me why this does not solve the > > problem. > > > > > > ___ > > 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] Where and How to Test the Strong Programs?
How does it deal with other gtp commands sent to it?Perhaps it can be used. Maybe Christoph can experiment with it. - Don Rémi Coulom wrote: > Don Dailey wrote: >> I saw that you made an illegal move! >> The way to do this is to the take the viewing client and hack it. >> Then you would get a nice gui and legal move testing (at the least the >> package to do legal move testing is there even if it's not being used.) >> >> If you are typing your moves in manually, you could at least pull the >> object oriented gogame package out and use it to verify the moves so you >> don't mistype an illegal move. It's dirt simple to use. >> >> I could probably do a gui since I plan to build a graphical engine >> client anyway.But I don't really want humans playing except as a >> special experiment. >> >> - Don > Note that the "gtpdisplay" tool that comes with gogui does this > already. You enter moves in the GUI, and they are sent as reply to the > "genmove" command. > > http://gogui.sourceforge.net/doc/reference-gtpdisplay.html > > Rémi > ___ > 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] Where and How to Test the Strong Programs?
I was wondering if gogui could be used - it would have to emulate a go program somehow. But gogui is a controller, not a program. However I know it comes with all kinds of filters to do various things. If it can be made to act like a go engine (where a human is the "brains") then it could be connected to the cgo3.tcl script directly. - Don Rémi Coulom wrote: > Don Dailey wrote: >> I saw that you made an illegal move! >> The way to do this is to the take the viewing client and hack it. >> Then you would get a nice gui and legal move testing (at the least the >> package to do legal move testing is there even if it's not being used.) >> >> If you are typing your moves in manually, you could at least pull the >> object oriented gogame package out and use it to verify the moves so you >> don't mistype an illegal move. It's dirt simple to use. >> >> I could probably do a gui since I plan to build a graphical engine >> client anyway.But I don't really want humans playing except as a >> special experiment. >> >> - Don > Note that the "gtpdisplay" tool that comes with gogui does this > already. You enter moves in the GUI, and they are sent as reply to the > "genmove" command. > > http://gogui.sourceforge.net/doc/reference-gtpdisplay.html > > Rémi > ___ > 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] Where and How to Test the Strong Programs?
Don Dailey wrote: I saw that you made an illegal move! The way to do this is to the take the viewing client and hack it. Then you would get a nice gui and legal move testing (at the least the package to do legal move testing is there even if it's not being used.) If you are typing your moves in manually, you could at least pull the object oriented gogame package out and use it to verify the moves so you don't mistype an illegal move. It's dirt simple to use. I could probably do a gui since I plan to build a graphical engine client anyway.But I don't really want humans playing except as a special experiment. - Don Note that the "gtpdisplay" tool that comes with gogui does this already. You enter moves in the GUI, and they are sent as reply to the "genmove" command. http://gogui.sourceforge.net/doc/reference-gtpdisplay.html Rémi ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
On Tue, 4 Dec 2007, Don Dailey wrote: But I don't really want humans playing except as a special experiment. I agree. But it's an interesting experiment ... Christoph ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
On Tue, 4 Dec 2007, Don Dailey wrote: It would be awkward at best. I could build a client to do this, but the human would have to be willing to sit and play games at the moment they were scheduled. You are right ... it's very awkward. I lost one game by typo and another by time. Christoph ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
I saw that you made an illegal move! The way to do this is to the take the viewing client and hack it. Then you would get a nice gui and legal move testing (at the least the package to do legal move testing is there even if it's not being used.) If you are typing your moves in manually, you could at least pull the object oriented gogame package out and use it to verify the moves so you don't mistype an illegal move. It's dirt simple to use. I could probably do a gui since I plan to build a graphical engine client anyway.But I don't really want humans playing except as a special experiment. - Don Christoph Birk wrote: > On Tue, 4 Dec 2007, Don Dailey wrote: >>> Yes, that would work. >>> Some humans also could play on CGOS (just for a while) to establish >>> a conversion from CGOS-ELO to human-ranks. >>> >> It would be awkward at best. I could build a client to do this, but >> the human would have to be willing to sit and play games at the moment >> they were scheduled. >> >> It would be better to get a handful of players to AGREE in advance to >> play particular programs on KGS but under very CGOS like conditions >> including the time control. They should know in advance they are >> part of the experiment and that they should not experiment but should >> play full strength and not abandon the games or take-back moves, etc. > > Yes, this would be better ... but will it happen? > I quickly hacked my program to allow me to make moves > and I am playing currently as test-3k (my AGA rating) > (so much for 'C' beeing slow for prototyping :-) > > Christoph > > ___ > 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] Elo and handicap matching
The only issue is that I don't know if GnuGo is representative of 19x19 to 9x9 go strength. I am interested in knowing how a human 19x19 scales down to 9x9 play. It's well known that programs scale up poorly. However, this data should still be quite useful. - Don Alain Baeckeroot wrote: > Le mardi 4 décembre 2007, Don Dailey a écrit : > > >> For 9x9 ELO works better. For 19x19 it's less clear cut.The >> handicap system appears to be a good system at 19x19 and has the very >> nice merit of allowing grossly mismatched players to compete. I >> think the two systems can be married by adding a fixed offset per stone >> handicap to your ELO. >> >> > > Stats from official european go federation, on 150 000 games nearly solves > the problem of doing ELO vs handicap matching. > http://gemma.ujf.cas.cz/~cieply/GO/statev.html > > Just need to find one anchor, lets says gnugo , rated 6k on kgs in 2007-11... > > Alain > > PS i posted this link some times on the list, but nobody seems to > consider it is useful (except Sylvain Gelly ;-) > I would be glad if someone could explain to me why this does not solve the > problem. > > > ___ > 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] Where and How to Test the Strong Programs?
On Tue, 4 Dec 2007, Don Dailey wrote: Yes, that would work. Some humans also could play on CGOS (just for a while) to establish a conversion from CGOS-ELO to human-ranks. It would be awkward at best. I could build a client to do this, but the human would have to be willing to sit and play games at the moment they were scheduled. It would be better to get a handful of players to AGREE in advance to play particular programs on KGS but under very CGOS like conditions including the time control. They should know in advance they are part of the experiment and that they should not experiment but should play full strength and not abandon the games or take-back moves, etc. Yes, this would be better ... but will it happen? I quickly hacked my program to allow me to make moves and I am playing currently as test-3k (my AGA rating) (so much for 'C' beeing slow for prototyping :-) Christoph ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Elo and handicap matching
On Tue, 4 Dec 2007, Alain Baeckeroot wrote: For 9x9 ELO works better. For 19x19 it's less clear cut.The handicap system appears to be a good system at 19x19 and has the very nice merit of allowing grossly mismatched players to compete. I think the two systems can be married by adding a fixed offset per stone handicap to your ELO. Stats from official european go federation, on 150 000 games nearly solves the problem of doing ELO vs handicap matching. http://gemma.ujf.cas.cz/~cieply/GO/statev.html Just need to find one anchor, lets says gnugo , rated 6k on kgs in 2007-11... That's nice, but on 19x19. I was interested in calibrating CGOS 9x9 versus some moderatly strong humans. Christoph ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Elo and handicap matching
Le mardi 4 décembre 2007, Don Dailey a écrit : > > > For 9x9 ELO works better. For 19x19 it's less clear cut.The > handicap system appears to be a good system at 19x19 and has the very > nice merit of allowing grossly mismatched players to compete. I > think the two systems can be married by adding a fixed offset per stone > handicap to your ELO. > Stats from official european go federation, on 150 000 games nearly solves the problem of doing ELO vs handicap matching. http://gemma.ujf.cas.cz/~cieply/GO/statev.html Just need to find one anchor, lets says gnugo , rated 6k on kgs in 2007-11... Alain PS i posted this link some times on the list, but nobody seems to consider it is useful (except Sylvain Gelly ;-) I would be glad if someone could explain to me why this does not solve the problem. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
Maybe it should be an official tournament on KGS. We should probably make it invitation only for bots and open to 1d+ from KGS. For invitation, maybe it should be 2200+ ELO bots? Looking at http://cgos.boardspace.net/9x9/standings.html, that seems to be: GreenPeep (2550) Zen (2472) MoGo (not listed, but obviously near the top) Goanna (2251) CrazyStone (2214, 2416?) Leela (2205) Valkyria (2202) If the threshold is 2400 ELO, the list may be more reasonable: GreenPeep (2550) Zen (2472) CrazyStone (2416?) MoGo (...) On 12/4/07, Don Dailey <[EMAIL PROTECTED]> wrote: > > > Christoph Birk wrote: > > On Tue, 4 Dec 2007, Chris Fant wrote: > >> MoGo. But it seems that it hasn't been playing recently (anyway, you > >> would have had no idea of the settings and hardware used). You could > >> play against it on your own hardware to understand it's strength > >> against a human, and let it get a CGOS rating using the same hardware > >> whenever you are not playing it. > > > > Yes, that would work. > > Some humans also could play on CGOS (just for a while) to establish > > a conversion from CGOS-ELO to human-ranks. > > > > Christoph > > It would be awkward at best. I could build a client to do this, but > the human would have to be willing to sit and play games at the moment > they were scheduled. > > It would be better to get a handful of players to AGREE in advance to > play particular programs on KGS but under very CGOS like conditions > including the time control. They should know in advance they are > part of the experiment and that they should not experiment but should > play full strength and not abandon the games or take-back moves, etc. > > - 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 mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] erm...
I would like to add that we are calibrating against 19x19 players. Even though their ratings are based on 19x19 play we just want a mapping from 19x19 dan to 9x9 cgos. The assumption is that 1 Dan 19x19 = 1 Dan 9x9 and on average this will be true. We don't expect to get a perfect table out of this - just a rough rule of thumb. I would mention that KGS in a sense does this. When you play 9x9 they still display your rating even if you believe there is no correlation or relevance.Of course they don't rate 9x9 games but you can still judge the quality of your opponent by looking at his 19x19 KGS ranking. - Don Don Dailey wrote: > But it does have real meaning. People talk about Dan level 9x9 go > programs and so all I'm looking for is a way to instrument this in a > meaningful way. > > If a 9x9 program is estimated to be 2 dan on CGOS, it means a typical 1 > dan player will lose to it and a typical 3 dan player will beat it. > It's a matter of adjusting the constants to make sure this is so. > > We cannot use the kyu/dan system for handicap on 9x9, it doesn't work > and that's understood - we are just trying to get a sense of how these > 9x9 programs compare to humans. > > - Don > > > > > ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
Christoph Birk wrote: > On Tue, 4 Dec 2007, Chris Fant wrote: >> MoGo. But it seems that it hasn't been playing recently (anyway, you >> would have had no idea of the settings and hardware used). You could >> play against it on your own hardware to understand it's strength >> against a human, and let it get a CGOS rating using the same hardware >> whenever you are not playing it. > > Yes, that would work. > Some humans also could play on CGOS (just for a while) to establish > a conversion from CGOS-ELO to human-ranks. > > Christoph It would be awkward at best. I could build a client to do this, but the human would have to be willing to sit and play games at the moment they were scheduled. It would be better to get a handful of players to AGREE in advance to play particular programs on KGS but under very CGOS like conditions including the time control. They should know in advance they are part of the experiment and that they should not experiment but should play full strength and not abandon the games or take-back moves, etc. - 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/
Re: [computer-go] erm...
But it does have real meaning. People talk about Dan level 9x9 go programs and so all I'm looking for is a way to instrument this in a meaningful way. If a 9x9 program is estimated to be 2 dan on CGOS, it means a typical 1 dan player will lose to it and a typical 3 dan player will beat it. It's a matter of adjusting the constants to make sure this is so. We cannot use the kyu/dan system for handicap on 9x9, it doesn't work and that's understood - we are just trying to get a sense of how these 9x9 programs compare to humans. - Don steve uurtamo wrote: > not to put too fine a point on it, but "estimating dan ranks" via > 9x9 games is a bit silly. it doesn't actually capture any extra > information about the program, since there's no such thing as > a 9x9 rank to compare with/against, much less a dan rank. > > ELO works well because it's strictly arbitrary relative to the > anchor. human kyu/dan ranks are arbitrary as well, but there > are thousands of people feeding ELO (or something like it) in > and out of the system, and there's an easily measured "top end". > > i'm just saying, it will look a bit silly -- imagine if when chess > programs started to completely dominate something like 5x5 chess, > the strongest programs were labeled "master level play". > > well, sure, at 5x5 chess... > > there's a reason that KGS won't assign a rank to a player that > only plays 9x9 games. the rating system isn't really built for > a board that small, in the sense that handicaps are massively > nonlinear at that size. the main motivation for kyu/dan rankings > is to establish handicap stones, after all... > > s. > > > > > > > Looking for last minute shopping deals? > Find them fast with Yahoo! Search. > http://tools.search.yahoo.com/newsearch/category.php?category=shopping > ___ > 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] Where and How to Test the Strong Programs?
On Tue, 4 Dec 2007, Chris Fant wrote: MoGo. But it seems that it hasn't been playing recently (anyway, you would have had no idea of the settings and hardware used). You could play against it on your own hardware to understand it's strength against a human, and let it get a CGOS rating using the same hardware whenever you are not playing it. Yes, that would work. Some humans also could play on CGOS (just for a while) to establish a conversion from CGOS-ELO to human-ranks. Christoph ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] erm...
On Tue, 4 Dec 2007, steve uurtamo wrote: not to put too fine a point on it, but "estimating dan ranks" via 9x9 games is a bit silly. it doesn't actually capture any extra information about the program, since there's no such thing as a 9x9 rank to compare with/against, much less a dan rank. I don't claim that this "9x9 CGOS rank" will be equal to ranks on 19x19. But it should work the other way round. If several "human ranked players" played on CGOS and established an ELO rating we would have a good estimate of how strong GGOS ratings are. Christoph ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
MoGo. But it seems that it hasn't been playing recently (anyway, you would have had no idea of the settings and hardware used). You could play against it on your own hardware to understand it's strength against a human, and let it get a CGOS rating using the same hardware whenever you are not playing it. On Dec 4, 2007 4:13 PM, Christoph Birk <[EMAIL PROTECTED]> wrote: > On Tue, 4 Dec 2007, Gunnar Farnebäck wrote: > > terry mcintyre wrote: > >> Some of the MonteGNU code was just released on CVS. Check out Gnugo's > >> development pages. > > > > Don't expect that code to do better than 2000 on CGOS though > > (mgtest2). The remaining code used by MonteGNU is still too messy. > > That's why I asked for 'MonteGNU'. It has a well established rating > on CGOS (23670 games). > I there any other (fairly) strong CGOS 9x9 program that's available > for download AND that has an reliable rating on CGOS? > > Christoph > > ___ > 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] Speed of generating random playouts
On 12/4/07, Álvaro Begué <[EMAIL PROTECTED]> wrote: > > How do you keep that updated cheaply as moves are made? For instance, if I > put a black stone next to a white chain, how do I update the OR and AND > pseudoliberty values for that chain? John's complicated solution only > requires storing a sum, which is always easy to update (in my example, > susbtract the value associated with the point where black just played). Oh, yes, I see. I didn't think of that. Seemed nice at first... ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] erm...
not to put too fine a point on it, but "estimating dan ranks" via 9x9 games is a bit silly. it doesn't actually capture any extra information about the program, since there's no such thing as a 9x9 rank to compare with/against, much less a dan rank. ELO works well because it's strictly arbitrary relative to the anchor. human kyu/dan ranks are arbitrary as well, but there are thousands of people feeding ELO (or something like it) in and out of the system, and there's an easily measured "top end". i'm just saying, it will look a bit silly -- imagine if when chess programs started to completely dominate something like 5x5 chess, the strongest programs were labeled "master level play". well, sure, at 5x5 chess... there's a reason that KGS won't assign a rank to a player that only plays 9x9 games. the rating system isn't really built for a board that small, in the sense that handicaps are massively nonlinear at that size. the main motivation for kyu/dan rankings is to establish handicap stones, after all... s. Looking for last minute shopping deals? Find them fast with Yahoo! Search. http://tools.search.yahoo.com/newsearch/category.php?category=shopping ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
On Tue, 4 Dec 2007, Gunnar Farnebäck wrote: terry mcintyre wrote: Some of the MonteGNU code was just released on CVS. Check out Gnugo's development pages. Don't expect that code to do better than 2000 on CGOS though (mgtest2). The remaining code used by MonteGNU is still too messy. That's why I asked for 'MonteGNU'. It has a well established rating on CGOS (23670 games). I there any other (fairly) strong CGOS 9x9 program that's available for download AND that has an reliable rating on CGOS? Christoph ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
On Tue, 4 Dec 2007, Don Dailey wrote: note: this is only to estimate the playing strength relative to a 19x19 player since there is no real system that makes sense for 9x9. I would simple put this on the crosstable web pages in parenthesis. e.g. Rated: 2410 (1.1d est.) I don't think this makes sense until it's calibrated. Christoph ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
terry mcintyre wrote: > Some of the MonteGNU code was just released on CVS. Check out Gnugo's > development pages. Don't expect that code to do better than 2000 on CGOS though (mgtest2). The remaining code used by MonteGNU is still too messy. /Gunnar ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
Let's make a wild guess.What if I made the web site report approximate strength using the following formula: dan = (elo - 2300) / 100 So a 2400 player is 1 dan, a 2500 player is 2 dan etc. Here is a table: 2300 - 1.0 kyu 2310 - 0.9 kyu 2320 - 0.8 kyu ... 2400 - 1.0 dan 2410 - 1.1 dan Does this scale and these constants seem reasonable? Am I off a class in either direction? note: this is only to estimate the playing strength relative to a 19x19 player since there is no real system that makes sense for 9x9. I would simple put this on the crosstable web pages in parenthesis. e.g. Rated: 2410 (1.1d est.) - Don Christoph Birk wrote: > Robert Jasiek wrote: >>> Where can one play the latest versions of MoGo or other, similarly >>> strong programs? > > Would it be possible to publish the MonteGNU code? > If yes, then a few dan-players could play each at least 20 games > against it and publish their results. That would allow for a > rough estimate of MonteGNU's strength (9x9 CGOS = 2100 ELO) > and a (first) normalization point for CGOS. > > Christoph > > ___ > 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] Speed of generating random playouts
On Dec 4, 2007 3:57 PM, Zach Wegner <[EMAIL PROTECTED]> wrote: > On Nov 13, 2007 2:44 PM, Jason House <[EMAIL PROTECTED]> wrote: > > > > > > > On Nov 13, 2007 3:32 PM, John Tromp <[EMAIL PROTECTED]> wrote: > > > > > > Is there any known way to get the best of the both worlds? :-) > > > > > > Yes, you can generalize pseudoliberties by extending them > > > with another field, such that if the (summed) pseudoliberty field > > > is between 1 and 4, then the other (summed) field will tell you if all > > > these > > > are coming from a single true liberty. > > > > > > Can you elaborate on this? > > > > I'm a bit late on this, and I'm not sure if this hasn't been suggested > yet. But, instead of these complicated codes, you could store in the extra > field both the AND and the OR of all pseudoliberties. Then if they are > equal, you have one real liberty. > > Example: > Pseudoliberties at 27, 17, and 9 (binary 11011, 10001, and 01001), > AND=1 and OR=11011. AND != OR so there's more than one pseudoliberty. > How do you keep that updated cheaply as moves are made? For instance, if I put a black stone next to a white chain, how do I update the OR and AND pseudoliberty values for that chain? John's complicated solution only requires storing a sum, which is always easy to update (in my example, susbtract the value associated with the point where black just played). Álvaro. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
Some of the MonteGNU code was just released on CVS. Check out Gnugo's development pages. Terry McIntyre <[EMAIL PROTECTED]> They mean to govern well; but they mean to govern. They promise to be kind masters; but they mean to be masters. -- Daniel Webster - Original Message From: Christoph Birk <[EMAIL PROTECTED]> To: computer-go Sent: Tuesday, December 4, 2007 12:14:24 PM Subject: Re: [computer-go] Where and How to Test the Strong Programs? Robert Jasiek wrote: >> Where can one play the latest versions of MoGo or other, similarly strong >> programs? Would it be possible to publish the MonteGNU code? If yes, then a few dan-players could play each at least 20 games against it and publish their results. That would allow for a rough estimate of MonteGNU's strength (9x9 CGOS = 2100 ELO) and a (first) normalization point for CGOS. Christoph ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ Looking for last minute shopping deals? Find them fast with Yahoo! Search. http://tools.search.yahoo.com/newsearch/category.php?category=shopping___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Speed of generating random playouts
On Nov 13, 2007 2:44 PM, Jason House <[EMAIL PROTECTED]> wrote: > > > On Nov 13, 2007 3:32 PM, John Tromp <[EMAIL PROTECTED]> wrote: > > > > Is there any known way to get the best of the both worlds? :-) > > > > Yes, you can generalize pseudoliberties by extending them > > with another field, such that if the (summed) pseudoliberty field > > is between 1 and 4, then the other (summed) field will tell you if all > > these > > are coming from a single true liberty. > > > Can you elaborate on this? > > I'm a bit late on this, and I'm not sure if this hasn't been suggested yet. But, instead of these complicated codes, you could store in the extra field both the AND and the OR of all pseudoliberties. Then if they are equal, you have one real liberty. Example: Pseudoliberties at 27, 17, and 9 (binary 11011, 10001, and 01001), AND=1 and OR=11011. AND != OR so there's more than one pseudoliberty. > > ___ > 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] The global search myth
Relatively speaking chess eval of adding piece values together and doing nothing else is far closer to optimal evaluation function that what is currently available in Go. A GOOD go evaluation function probably needs to incorporate lookahead... Through most of the game, the difference between "a dead group" and "a live group" can only be distinguished by answering "What happens here if I do THIS?" And to make that investigation supportable, you need a good move-evaluator. A pattern-recognizer with a taste for 'good shape.' In other words, all these functions are interdependent and virtually need one another just to "define" (in an operational sense) what they need to consider! Forrest Curo This message was sent using IMP, the Internet Messaging Program. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
Rémi Coulom wrote: Hi, 13x13 StoneCrazy is currently connected to CGOS (computer go room). It will stay there for about 24h. Rémi So far, it lost 1 game against 3d, and 2 games against 2d. In this game, it started a nice ko fight at move 69 (but lost): http://files.gokgs.com/games/2007/12/4/BadRobot-StoneCrazy.sgf Semeais in the corner with a nakade generate huge catastrophes. Unless I fix this, it will still lose to 5k players when they occur, whatever computational power I use. I have also connected Crazy Stone on a core 2 duo to 9x9 CGOS, and will leave it there for a while. Rémi ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
Robert Jasiek wrote: Where can one play the latest versions of MoGo or other, similarly strong programs? Would it be possible to publish the MonteGNU code? If yes, then a few dan-players could play each at least 20 games against it and publish their results. That would allow for a rough estimate of MonteGNU's strength (9x9 CGOS = 2100 ELO) and a (first) normalization point for CGOS. Christoph ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
At the Cotsen Open the encouragement is a prize for the best program. It has not been very satisfying for me to have SlugGo win it the past two years by the default of being the only program present. I would be much happier to have others show up too. I have heard from one programmer who says he will enter a program next year. In the past the prize was cash, but at my request they changed it to a trophy this year ... money disappears but trophies last. If others are going to enter programs next year I encourage them to tell the organizers which is more valuable to them. Cheers, David On 4, Dec 2007, at 10:27 AM, Robert Jasiek wrote: David Doshay wrote: When tournament organizers allow and encourage it! Some (local) European tournaments would allow it. (Some have already done it.) "Encourage" - not yet :) -- robert jasiek ___ 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] New engine? From a Chess programmer perspective.
On Tue, 4 Dec 2007, David Fotland wrote: It's not clear if you are talking about professional Dan level or Amateur Dan level. I've played the top 9x9 programs at 9x9, and so have several other amateur Dan players, and I think we all agree that the top 9x9 programs have reached amateur Dan level. I don't think these programs are as strong as professional Dan players. Yes, they are amateur dan level, but not (yet) "high" (5+) dan. Chrsitoph ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
David Doshay wrote: When tournament organizers allow and encourage it! Some (local) European tournaments would allow it. (Some have already done it.) "Encourage" - not yet :) -- robert jasiek ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] New engine? From a Chess programmer perspective.
David Fotland wrote: It's not clear if you are talking about professional Dan level or Amateur Dan level. I have meant the latter. I've played the top 9x9 programs at 9x9, and so have several other amateur Dan players, and I think we all agree that the top 9x9 programs have reached amateur Dan level. Is there some summary of those tests, which is more profound than anecdotes? -- robert jasiek ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] New engine? From a Chess programmer perspective.
Don Dailey wrote: > For 19x19 it's less clear cut.The handicap system appears to be a good system I can't agree, but this has already been discussed at rec.games.go. (E.g., a player does not need to be good at handicap go if is good in even games.) -- robert jasiek ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
On 4, Dec 2007, at 3:59 AM, Robert Jasiek wrote: When will we see the strong programs entering real world tournaments? When tournament organizers allow and encourage it! At this time AGA rules are that games against computers are not counted in a human player's ranking. The Cotsen Open allows programs to enter, so I am there every year, with SlugGo presently entered as a 10 kyu on a full board. SlugGo has won against 8 Kyu and lost to 12 kyu players. I am expecting that there will be something happening at the US Go Congress 2008 in Portland. Cheers, David ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] New engine? From a Chess programmer perspective.
David Fotland wrote: > You can't add a fixed ELO offset per stone because games between stronger > players have much lower variance in score. A handicap stone is > approximately a score offset (about 7.5 points for the first handicap stone, > and about 15 points for each additional stone). > > ELO measures probability of winning. A 2 stone handicap game between equal > high dan players has a very high probability of black winning, so there must > be a large ELO offset. A 2 stone game between low kyu players gives black a > small increment in winning probability, so the ELO offset must be small. > > David > Ok, so that means we would have to have a more sophisticated formula. The formula could be dynamically adjusted over time by the server itself to reflect actual results more accurately. - Don >> I think the two systems can be married by adding a fixed offset per stone >> handicap to your ELO. >> >> - 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/
Re: [computer-go] Where and How to Test the Strong Programs?
Robert Jasiek wrote: > Where can one play the latest versions of MoGo or other, similarly > strong programs? It is said that some programs are on KGS, but I > cannot find them. How to find them? Is it possible to play against > them as a human on CGOS? I, German 5d, would want to play even games > on 19x19, 13x13, or 9x9 to learn more about their current playing > level. How would that be used for assessing the program's strength? > E.g., KGS ratings are inaccurate by +-600; the standard deviation of > an EGF high dan rating I subjectively estimate as roughly +-70. So IMO > ratings are not useful to assess a program's strength against humans. > When will we see the strong programs entering real world tournaments? > It would be nice to "normalize" the ratings of CGOS to human terms. If we had a great deal of data points, we could construct an ELO to kyu/dan formula which could be posted along with the ELO ratings on the web page. For instance here is a stab: dan = (elo - 2400) / 100 The issue is which constants to use and is this form accurate enough to encompass a fairly large range of rankings? Can it be reasonably accurate at 15k and still accurate at 2 Dan if the constants are correctly chosen or does it need to be more sophisticated? At any rate, I'm more interested in accuracy near the levels of the top programs. Since we are comparing to human players and also comparing 19x19 rankings to 9x9 rankings, this is a little dicey - but it will still work.What we want to know is how a given program is likely to do against a particular human player and this is just an estimate.The human player is of course assumed to be a 19x19 rated player. The other problem is that there is no universal calibration - a 1 Dan KGS player may be 2 Dan in some other system.But perhaps KGS is universal enough for our needs. We can rate any number of players on KGS by just playing them for a hundred games or so - but a lot of games are aborted, or time added or take-backs are made, etc. Also, the games should be played at the same time control as on CGOS. - Don ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] New engine? From a Chess programmer perspective.
You can't add a fixed ELO offset per stone because games between stronger players have much lower variance in score. A handicap stone is approximately a score offset (about 7.5 points for the first handicap stone, and about 15 points for each additional stone). ELO measures probability of winning. A 2 stone handicap game between equal high dan players has a very high probability of black winning, so there must be a large ELO offset. A 2 stone game between low kyu players gives black a small increment in winning probability, so the ELO offset must be small. David > I think the two systems can be married by adding a fixed offset per stone > handicap to your ELO. > > - Don ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] New engine? From a Chess programmer perspective.
It's not clear if you are talking about professional Dan level or Amateur Dan level. I've played the top 9x9 programs at 9x9, and so have several other amateur Dan players, and I think we all agree that the top 9x9 programs have reached amateur Dan level. I don't think these programs are as strong as professional Dan players. At 19x19 their strength is much less clear. I don't think they are quite to amateur Dan yet. David Fotland > -Original Message- > From: [EMAIL PROTECTED] [mailto:computer-go- > [EMAIL PROTECTED] On Behalf Of Robert Jasiek > Sent: Tuesday, December 04, 2007 7:54 AM > To: computer-go > Subject: Re: [computer-go] New engine? From a Chess programmer > perspective. > > Don Dailey wrote: > > Just a few years ago it was widely held that computers will not reach > > Dan level "in my lifetime" even in 9x9 Go.When it happened in 9x9 > > go, it was not accepted - the day it happened passed us by and > nobody > > noticed it. It's probably still not common knowledge and it will > > take time for it to be generally believed. > > What is the basis of the claim that a program has reached a certain > human rank level? > > There should be systematic tests. Let it play against many humans. Let > it enter human tournaments. Use a meaningful evaluation context. > > For some such evaluation, let me refer to a useful handicap system for > 9x9, which has been used in some European 9x9 "Championships": For the > first 10 rank differences (0, 1,.., 9) decrease the komi from 6.5 for > an > even game by 1 point per extra rank. (Komi can become negative.) - OC, > I > prefer to see even games. OTOH, until the program rank is well known, > it may be suitable to let a simgle human (the more humans the better) > play until the handicap becomes stable. > > -- > robert jasiek > ___ > 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] New engine? From a Chess programmer perspective.
Robert Jasiek wrote: > Don Dailey wrote: >> Just a few years ago it was widely held that computers will not reach >> Dan level "in my lifetime" even in 9x9 Go.When it happened in 9x9 >> go, it was not accepted - the day it happened passed us by and nobody >> noticed it. It's probably still not common knowledge and it will >> take time for it to be generally believed. > > What is the basis of the claim that a program has reached a certain > human rank level? > > There should be systematic tests. Let it play against many humans. Let > it enter human tournaments. Use a meaningful evaluation context. Yes, there should be. Right now it's very informal and anecdotal. Strictly speaking it is not possible to evaluate any player - the ELO system is subject to the laws of probability.So no matter how many games you play people can claim it's a fluke and blame the playing conditions or other factors. That's the point I was really trying to make.People will be resistant to the idea no matter what - so the best you can do is supply overwhelming empirical evidence and then if some want to be unreasonable they are easily identified - at least by reasonable people. > > For some such evaluation, let me refer to a useful handicap system for > 9x9, which has been used in some European 9x9 "Championships": For the > first 10 rank differences (0, 1,.., 9) decrease the komi from 6.5 for > an even game by 1 point per extra rank. (Komi can become negative.) - > OC, I prefer to see even games. OTOH, until the program rank is well > known, it may be suitable to let a simgle human (the more humans the > better) play until the handicap becomes stable. > For 9x9 ELO works better. For 19x19 it's less clear cut.The handicap system appears to be a good system at 19x19 and has the very nice merit of allowing grossly mismatched players to compete. I think the two systems can be married by adding a fixed offset per stone handicap to your ELO. - Don ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
Robert Jasiek wrote: Where can one play the latest versions of MoGo or other, similarly strong programs? It is said that some programs are on KGS, but I cannot find them. How to find them? Is it possible to play against them as a human on CGOS? I, German 5d, would want to play even games on 19x19, 13x13, or 9x9 to learn more about their current playing level. How would that be used for assessing the program's strength? E.g., KGS ratings are inaccurate by +-600; the standard deviation of an EGF high dan rating I subjectively estimate as roughly +-70. So IMO ratings are not useful to assess a program's strength against humans. When will we see the strong programs entering real world tournaments? Hi, 13x13 StoneCrazy is currently connected to CGOS (computer go room). It will stay there for about 24h. Rémi ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] The global search myth
Relatively speaking chess eval of adding piece values together and > > doing nothing else is far closer to optimal evaluation function that > what is currently available in Go. > > Adding piece values together is just a crude starting point in chess. > It is true that there isn't an obvious candidate in go to play a > similar role, but it's not like you can get a grandmaster-level > program by counting beans. I think there is an obvious candidate - count stones and total liberties. It's not clear to me that it is worse than just counting material in chess - because that is a terrible evaluation function in chess just as counting stones and liberties is terrible in go. Go is more complex but stone count and territory are the most basic building blocks of an evaluation function just as head count is in chess. In chess a good evaluation function has to measure how good each piece really is. A trapped rook or bad pawn is like a dead group in go.In a really good go evaluation function you have to measure the probability of a groups survival and the "goodness" of each stone. In chess you must do the same. In go it's more clear cut what the value of a stone is. It's possible to see that many stones together are "dead." In chess it's rare to see this except by search.So I don't see any difference - I think these two respective evaluation functions are roughly equivalent and equally bad. Of course go is a more complex game and the horizon effect is far more deadly in GO than in Chess so this evaluation produces an even weaker Go program than in Chess. I'm not sure what is meant by "farther from optimal" because the tree search is far more complex and horizon effects are greater. But chess can have pretty wicked horizon effects too and thus go presents more difficult version of the same problems - not different problems.(Yes, the solution in chess is often to ignore problems - but that is not the same as saying they are not problems.) - Don > > Álvaro. > > > > > > ___ > 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] Where and How to Test the Strong Programs?
On 12/4/07, Chris Fant <[EMAIL PROTECTED]> wrote: > > What I consider more of an issue is that MoGo seems to be very > > sensitive to (undocumented) configuration options. Such issues > > probably exist with all engines. It'd probably be smarter to set up a > > day where strong bots would connect to CGOS and invite dan-level > > players to challenge them. > > You mean KGS, right? I don't think humans on CGOS are an appropriate > direction for that server. Yes, KGS, sorry. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
> What I consider more of an issue is that MoGo seems to be very > sensitive to (undocumented) configuration options. Such issues > probably exist with all engines. It'd probably be smarter to set up a > day where strong bots would connect to CGOS and invite dan-level > players to challenge them. You mean KGS, right? I don't think humans on CGOS are an appropriate direction for that server. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] New engine? From a Chess programmer perspective.
Don Dailey wrote: Just a few years ago it was widely held that computers will not reach Dan level "in my lifetime" even in 9x9 Go.When it happened in 9x9 go, it was not accepted - the day it happened passed us by and nobody noticed it. It's probably still not common knowledge and it will take time for it to be generally believed. What is the basis of the claim that a program has reached a certain human rank level? There should be systematic tests. Let it play against many humans. Let it enter human tournaments. Use a meaningful evaluation context. For some such evaluation, let me refer to a useful handicap system for 9x9, which has been used in some European 9x9 "Championships": For the first 10 rank differences (0, 1,.., 9) decrease the komi from 6.5 for an even game by 1 point per extra rank. (Komi can become negative.) - OC, I prefer to see even games. OTOH, until the program rank is well known, it may be suitable to let a simgle human (the more humans the better) play until the handicap becomes stable. -- robert jasiek ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
On 12/4/07, Chris Fant <[EMAIL PROTECTED]> wrote: > > > Where can one play the latest versions of MoGo or other, similarly > > > strong programs? > > > > But Mogo is now a free program.You can get a copy, find some good > > hardware and play at 9x9 and 19x19. > > > > But the released version is probably not the latest. Most strong bots will probably re-release when significant enhancements have been made. I think recent MoGo updates have been to support enhanced hardware (AKA, use MPI). On a normal computer, I'd assume it plays at about the same level. What I consider more of an issue is that MoGo seems to be very sensitive to (undocumented) configuration options. Such issues probably exist with all engines. It'd probably be smarter to set up a day where strong bots would connect to CGOS and invite dan-level players to challenge them. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
> > Where can one play the latest versions of MoGo or other, similarly > > strong programs? It is said that some programs are on KGS, but I > > cannot find them. How to find them? Is it possible to play against > > them as a human on CGOS? > CGOS is designed for computer/computer only.You could modify the > client to accept moves or build a client and pretend you are a bot - but > you have no control over the scheduling algorithm - you would be forced > to accept whatever pairing and color was assigned to you. > > But Mogo is now a free program.You can get a copy, find some good > hardware and play at 9x9 and 19x19. > But the released version is probably not the latest. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Where and How to Test the Strong Programs?
Robert Jasiek wrote: > Where can one play the latest versions of MoGo or other, similarly > strong programs? It is said that some programs are on KGS, but I > cannot find them. How to find them? Is it possible to play against > them as a human on CGOS? CGOS is designed for computer/computer only.You could modify the client to accept moves or build a client and pretend you are a bot - but you have no control over the scheduling algorithm - you would be forced to accept whatever pairing and color was assigned to you. But Mogo is now a free program.You can get a copy, find some good hardware and play at 9x9 and 19x19. - Don > I, German 5d, would want to play even games on 19x19, 13x13, or 9x9 to > learn more about their current playing level. How would that be used > for assessing the program's strength? E.g., KGS ratings are inaccurate > by +-600; the standard deviation of an EGF high dan rating I > subjectively estimate as roughly +-70. So IMO ratings are not useful > to assess a program's strength against humans. When will we see the > strong programs entering real world tournaments? > ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] New engine? From a Chess programmer perspective.
Hi Gian-Carlo, There is an interesting phenomenon going on when it comes to the perception and advertisement of game playing strength. One is that people take time to accept concepts they are used to thinking differently about. I remember one human (chess) player who was pretty weak for many years, then all of a sudden he "got serious" and gains hundreds of ELO points in a couple of years - becoming a master. Many players could not accept this and it took a few years (even though he continued to play actively) for people to stop believing he was grossly overrated and be accepted as a strong player. Just a few years ago it was widely held that computers will not reach Dan level "in my lifetime" even in 9x9 Go.When it happened in 9x9 go, it was not accepted - the day it happened passed us by and nobody noticed it. It's probably still not common knowledge and it will take time for it to be generally believed. Another phenomenon, is what I call the contempt factor. In the old days of computer chess, you would buy a program or machine that was certified to be a certain strength level. But once you got comfortable and familiar with the program, you started learning it's weaknesses and got time to witness a few stupid moves - your contempt of it grew and your estimation of it's strength diminished. The same thing happens with human players but for some reason we don't hold that against them. It's not that the program is not as strong as advertised, it's just that the magic goes away once we get to take it apart and see how it works. I don't know if your program is high Dan level or not.I suspect it's "low Dan" level if I had to guess but it could be tested with a formalized match on KGS. I suggest multiple players for variety - a single player match is not a good test. - Don Gian-Carlo Pascutto wrote: >> On Sun, 2 Dec 2007, Russell Wallace wrote: >> >>> I haven't seen Leela before, but the claim of high dan-level >>> performance on 9x9 is certainly interesting. >>> >> I don't think 2200 ELO on the 9x9 CGOS is equivalent to 'high dan-level' >> play. >> > > I was under the impression that MoGo (approx 2350 CGOS) was > starting to cause trouble for pro players on 9 x 9. The released > Leela version is a bit stronger than the last on CGOS and uses > all CPUs, so "high dan level" was supposed to be a reasonable estimate. > > If someone has factual data[*] about 9 x 9 performance of > current bots I'll gladly revise the estimate on the webpage > on my own. > > [*] factual data is not: "I feel it's about 1kyu". Or "I played > a few games and it sucks in life & death. I had to take back a move > because I wasn't really concentrated but I beat it easily. Must be > less than 2 dan strength". > > One of the best things I found was a report from a 6 dan that > he won a match 5-2 against an older version of MoGo. That puts > MoGo at about 4 to 5 dan. I don't think what I said is > unreasonable, unless a 5 dan is not considered "high dan level". > > Arguing about this feels like a waste of time anyway. At the last > KGS tournament people were arguing that Crazy Stone is overrated > because "it can't be 1k". > > The last time I saw this was when "dan" was called "grandmaster". > > ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Great day for CrazyStone!
The summary looks good to me. Just to clarify HouseBot's round 3 performance... HouseBot would normally resign lost games, but it has a check in there that prevents resignation when it has not thought deeply enough about every move. 19x19 is such a big board that it does not hit that threshold in before endgame. In endgame, however, it contains logic to stop thinking early when there's nothing to think about. Unfortunately, the interaction of the two pieces of logic means that it does not resign in endgame. I guess it can be thought of as a stress test for the time management in other bots ;) It's time management code worked as expected. Doing some quick number crunching, I believe the time management code would want about 4.75 seconds left at such a late point in the game. Relatively speaking, 6 seconds is viewed as a lot of spare time. I was actually more worried about round 1 when HouseBot crashed at the start of the cleanup phase with 24 seconds left. I had to quickly restart the bot. That crash was because HouseBot had an internal buffer overflow as it queued up commands as kgsGtp replayed the entire game. The 458-move game overflowed the 256 element command buffer. Thankfully that didn't happen again when I restarted the bot. I increased the buffer size between rounds 1 and 2 to prevent that from happening again. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] The global search myth
Petri Pitkanen wrote: > I think Monte-Carlo is more attempting solve a different issue > altogether. Sure it is a search tree buyt main problem is the > evaluation function. Currently we do not know any good way to evaluate > the situation on go board until the game is at very late stages. And I > think - not that I could support this with any testing - that most of > the current evaluation function would not play better if they had > deeper global search and actually may play worse with wider global > search. > Yes, evaluation is a difficult problem in Go. UCT neatly "solves" that problem because UCT does not require a formal evaluation function. > Relatively speaking chess eval of adding piece values together and > doing nothing else is far closer to optimal evaluation function that > what is currently available in Go. I think the go equivalent is to add up the number of stones and the total number of liberties for each side, where a liberty only gets to count 1 time. This evaluation function has the characteristic that it returns the correct score when the game is complete if you use logical rules. Although some evaluation functions can be constructed which play worse with depth, that is not the general case.In general, any "reasonable" (not good - just reasonable) evaluation function will play better and better with increasing search depth. For instance the evaluation function I just suggested improves with depth. Even with no quies search of any kind - just using the raw evaluation above you can extrapolate and see that this must be true. Obviously, since it knows nothing about dead groups it's quite naive and plays very poorly at 1 ply. However, even at this depth it won't play on the edges of the board which makes it better than random already.At 1 ply it will be just barely smart enough to make a capture (which also makes it better than random.) At 2 ply it will be smart enough to defend against atari. It will get this wrong a lot, for instance it will try to defend groups that cannot be defended. It would probably also play out all ladders to the end from the defenders side - win or lose at 2 ply. But after each additional ply, it will be a little wiser than before, capable of something it was not capable of before. It will also refine what it already "knows", such as being a bit smarter about ladders. For instance at 4 ply, it will know whether to defend a 2 move ladder or whether it's futile, thereby using it's moves elsewhere. If you imagine such a search doing 30 or 40 ply (on a 9x9 board) it might actually be playing pretty strong.It would certainly play endings very well at this point. At low depth levels in the opening, such a program would scatter the stones randomly, trying to get 1 space and 4 liberties for each stone. Although it wouldn't play on the edge or corner for no reason, it would likely play a lot of second rank moves as you mention. There is some depth level, however, where this behavior will start to change. At some point it will learn that moves like b2 runs you out of options later. The opponent gains time attacking this stone from the outside and whether you are able to defend it or not, the opponent has placed a lot of stones where they count.On 9x9 boards I think it would start playing what looks like fairly normal moves at around 15-30 ply. At some point, probably well before it can perform an exhaustive search of the whole game, it will play very strongly, perhaps as good as the best players. I'm not claiming this is a "good" way to proceed, I'm just trying to refute the idea that it would play worse with depth - this is clearly not true.Of course it's possible to refine this evaluation considerably - it's pretty lousy as I described it but I used it as an example because it's easy to understand and describe. - Don > Also there is not much published information evaluation functions in > Go. Obviously a go programming is a business and giving out such > information does not make sense. Best publicly available thingy is > GnuGo and it does not even have one. > > Any simplistic Go-veal would probably result in very bad choices in > early stages of game - like playing on second row without proper > reason. So selective search is part of eval. And this shortcoming is > pretty obvious in MC programs. when they play on full sized board they > make extremely funny moves and so good result against humans only in > ultra-blitz conditions - humans scale better I guess. > > Petri > ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] New engine? From a Chess programmer perspective.
> Mogo is around 2500 on CGOS: > http://cgos.boardspace.net/9x9/cross/MoGo_psg7.html > This implies you believe the ratings didn't shift over time. http://computer-go.org/pipermail/computer-go/2007-October/011405.html http://cgos.boardspace.net/9x9/cross/MoGo_monothreadC.html http://cgos.boardspace.net/9x9/cross/goala1.html The MoGo team has worked for 5 months and gained...-200 ELO. http://cgos.boardspace.net/9x9/cross/greenpeep0.3.4.html http://cgos.boardspace.net/9x9/cross/greenpeep0.4.2.html Same phenomenon. > In Amsterdam, ajahuang (kgs 6d) played a few games against MoGo on 9x9, > and won them all. This can be seen in his history on KGS. That's a good data point which would drop the estimate a few ranks. -- GCP ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] The global search myth
On Dec 4, 2007 1:42 AM, Petri Pitkanen <[EMAIL PROTECTED]> wrote: > > There is something that the latest Monte Carlo programs have in common > > with the best chess programs - and seems to be the right way to > > structure a game tree search.Your selectivity should be > > progressive. In order to do this correctly you must re-visit nodes > > many times. Chess programs do it iteratively and Monte Carlo UCT type > > programs do it "best first" fashion. So the decision to prune any given > > move is a decision that is considered many times in the course of a > > search - each time taking advantage of additional information. > > > > I think Monte-Carlo is more attempting solve a different issue > altogether. Sure it is a search tree buyt main problem is the > evaluation function. Currently we do not know any good way to evaluate > the situation on go board until the game is at very late stages. And I > think - not that I could support this with any testing - that most of > the current evaluation function would not play better if they had > deeper global search and actually may play worse with wider global > search. > The way I see it, Monte-Carlo *is* an evaluation function. Run a few thousand simulations from a given position and the fraction of them that you win is an indication of how good the position was to begin with. I can't imagine why you would say that deeper/wider global search would hurt. The only reason not to search deeper or wider (I assume by that you mean "less selectively") is lack of resources. Relatively speaking chess eval of adding piece values together and > doing nothing else is far closer to optimal evaluation function that > what is currently available in Go. Adding piece values together is just a crude starting point in chess. It is true that there isn't an obvious candidate in go to play a similar role, but it's not like you can get a grandmaster-level program by counting beans. > Also there is not much published information evaluation functions in > Go. Obviously a go programming is a business and giving out such > information does not make sense. Best publicly available thingy is > GnuGo and it does not even have one. This was true of chess too, and there is more money to be made in computer chess, so my guess is that people are more secretive there. Any simplistic Go-veal would probably result in very bad choices in > early stages of game - like playing on second row without proper > reason. Any simplistic evaluation function in pretty much any game will result in horribly bad moves when there are no obvious tactical gains to be had. Try a closed position in chess, for example. > So selective search is part of eval. What? That doesn't follow at all. I'm not even sure what it means. > And this shortcoming is > pretty obvious in MC programs. when they play on full sized board they > make extremely funny moves and so good result against humans only in > ultra-blitz conditions - humans scale better I guess. Poor performance of current programs doesn't mean anything. People in the 80s could have made similar arguments saying that alpha-beta searchers with a simple evaluation function that considers material and positional values only cannot possibly be the right approach, as evidenced by their funny moves and their poor results against humans. It turns out that an evolution of that approach on much faster hardware plays better than humans. Álvaro. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] Where and How to Test the Strong Programs?
Where can one play the latest versions of MoGo or other, similarly strong programs? It is said that some programs are on KGS, but I cannot find them. How to find them? Is it possible to play against them as a human on CGOS? I, German 5d, would want to play even games on 19x19, 13x13, or 9x9 to learn more about their current playing level. How would that be used for assessing the program's strength? E.g., KGS ratings are inaccurate by +-600; the standard deviation of an EGF high dan rating I subjectively estimate as roughly +-70. So IMO ratings are not useful to assess a program's strength against humans. When will we see the strong programs entering real world tournaments? -- robert jasiek ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] New engine? From a Chess programmer perspective.
Christoph Birk wrote: I don't think 2200 ELO on the 9x9 CGOS is equivalent to 'high dan-level' play. Neither do I. In fact the whole kyu/dan rating system applies only to 19x19. 9x9 is not deep enough to have have so many ranks. On a 9x9 board an average amateur beats a pro with handicap 3. That amateur would be crushed by the pro with handicap 9 in 19x19. Jacques. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] New engine? From a Chess programmer perspective.
Gian-Carlo Pascutto wrote: If someone has factual data[*] about 9 x 9 performance of current bots I'll gladly revise the estimate on the webpage on my own. Mogo is around 2500 on CGOS: http://cgos.boardspace.net/9x9/cross/MoGo_psg7.html In Amsterdam, ajahuang (kgs 6d) played a few games against MoGo on 9x9, and won them all. This can be seen in his history on KGS. Rémi ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] New engine? From a Chess programmer perspective.
Hi all, Gian-Carlo Pascutto: <[EMAIL PROTECTED]>: >> On Sun, 2 Dec 2007, Russell Wallace wrote: >>> I haven't seen Leela before, but the claim of high dan-level >>> performance on 9x9 is certainly interesting. >> >> I don't think 2200 ELO on the 9x9 CGOS is equivalent to 'high dan-level' >> play. > >I was under the impression that MoGo (approx 2350 CGOS) was >starting to cause trouble for pro players on 9 x 9. The released >Leela version is a bit stronger than the last on CGOS and uses >all CPUs, so "high dan level" was supposed to be a reasonable estimate. MoGo's highest rating on cgos 9x9 was 2556. http://cgos.boardspace.net/9x9/cross/MoGo_G3.4.html -Hideki >If someone has factual data[*] about 9 x 9 performance of >current bots I'll gladly revise the estimate on the webpage >on my own. > >[*] factual data is not: "I feel it's about 1kyu". Or "I played >a few games and it sucks in life & death. I had to take back a move >because I wasn't really concentrated but I beat it easily. Must be >less than 2 dan strength". > >One of the best things I found was a report from a 6 dan that >he won a match 5-2 against an older version of MoGo. That puts >MoGo at about 4 to 5 dan. I don't think what I said is >unreasonable, unless a 5 dan is not considered "high dan level". > >Arguing about this feels like a waste of time anyway. At the last >KGS tournament people were arguing that Crazy Stone is overrated >because "it can't be 1k". > >The last time I saw this was when "dan" was called "grandmaster". -- [EMAIL PROTECTED] (Kato) ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] New engine? From a Chess programmer perspective.
> On Sun, 2 Dec 2007, Russell Wallace wrote: >> I haven't seen Leela before, but the claim of high dan-level >> performance on 9x9 is certainly interesting. > > I don't think 2200 ELO on the 9x9 CGOS is equivalent to 'high dan-level' > play. I was under the impression that MoGo (approx 2350 CGOS) was starting to cause trouble for pro players on 9 x 9. The released Leela version is a bit stronger than the last on CGOS and uses all CPUs, so "high dan level" was supposed to be a reasonable estimate. If someone has factual data[*] about 9 x 9 performance of current bots I'll gladly revise the estimate on the webpage on my own. [*] factual data is not: "I feel it's about 1kyu". Or "I played a few games and it sucks in life & death. I had to take back a move because I wasn't really concentrated but I beat it easily. Must be less than 2 dan strength". One of the best things I found was a report from a 6 dan that he won a match 5-2 against an older version of MoGo. That puts MoGo at about 4 to 5 dan. I don't think what I said is unreasonable, unless a 5 dan is not considered "high dan level". Arguing about this feels like a waste of time anyway. At the last KGS tournament people were arguing that Crazy Stone is overrated because "it can't be 1k". The last time I saw this was when "dan" was called "grandmaster". -- GCP ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] UEC cup result
UEC cup was held in 2007/12/01-02 in The University of Electro-communications, Tyofu, Tokyo, Japan. 27 programs competed. (GNU Go also entered as a guest, so all 28 programs) First day was Swiss 5R, and second day was tournament by best 16. And Crazy Stone won. 2nd was Katshunari and 3rd was MoGo. Crazy Stone played exhibition game against human 5dan player with no handicap game. Crazy Stone lost (professinal Go player Tei Meikou stopped the game halfway.) but it was interesting game. (SGF file will be available official page.) Rank Program Author 1st Crazy Stone Remi Coulom(France) (Invited program) 2nd KatsunariShinichi Sei 3rd MoGo Sylvain Gelly (France) (Invited program) 4th Aya Hiroshi Yamashita 5th GGMC Go Hideki Katoh 6th CarenKatsumi Kobayashi 7th ShikouSakugo Morihiko Tajima 8th Kinoa IgoGenki Yamada 9th RunGoHideaki Takahashi 10th GNU_ark Tsujii_labo(Nobuo Araki) 11th MayoiGo Masaki Murayama 12th Boozer Chihiro Hashimoto 13th martha Ichiro Ujiie 14th GOGATAKI Shigetaka Hisatomi 15th Igoppi Kazuhiko Ariyoshi 16th boon Kenta Sasaki I put some photos. http://www.yss-aya.com/photo/20071202uec/uec2007.html Swiss 5R result (first day) http://www.yss-aya.com/photo/20071202uec/uec1201/Htmls/PICT2139.html tournament table (second day) http://www.yss-aya.com/photo/20071202uec/uec1202/Htmls/PICT2176.html Detail result will be on official page.(not yet) http://jsb.cs.uec.ac.jp/~igo/eng/ The competion introduction on Japanese newspaper (in Japanese) http://mainichi.jp/enta/igo/news/20071201ke040014000c.html Regards, Hiroshi Yamashita ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/