Re: [Computer-go] oakfoam/Zen9 game
(2012/02/14 18:51), Nick Wedd wrote: On 13/02/2012 10:36, Yamato wrote: A similar thing happened in past tournaments, and we sent the log file to wms that time. I guess he doesn't fix this bug yet. I have now had a reply from wms. He wrote > Both of these players were on version 3.5.0 of the GTP client. This > is a fixed bug. If Zen9 had had version 3.5.3 or later, then it would > not have lost. So, everyone, please download and use the latest version of kgsGtp.jar. It is at http://www.gokgs.com/download.jsp OK, I didn't think this was a client problem. -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] oakfoam/Zen9 game
A similar thing happened in past tournaments, and we sent the log file to wms that time. I guess he doesn't fix this bug yet. -- Yamato (2012/02/13 19:26), Erik van der Werf wrote: Why guess? Yamato, don't you or Hideki keep log files with the output of kgsgtp? There should be some lines like: INFO: Disagreement over tournament scoring. Switching to cleanup mode. ... FINEST: Command sent to engine: undo ... FINEST: Command queued for sending to engine: kgs-genmove_cleanup b ... Erik On Mon, Feb 13, 2012 at 10:53 AM, Nick Wedd wrote: On 13/02/2012 00:10, Yamato wrote: I guess the server did not send the clean-up command to Zen. This seems unlikely. The server usually sends the clean-up command to both players. My guess is: the server sent the clean-up command to both players oakfoam passed the server took this pass, with the previous pass that was part of the game, to be two consecutive passes, and therefore the end of the clean-up phase. I have written to wms about this. Nick -- Nick Wedd n...@maproom.co.uk ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] oakfoam/Zen9 game
I guess the server did not send the clean-up command to Zen. -- Yamato (2012/02/12 22:23), Nick Wedd wrote: In my report http://www.weddslist.com/kgs/past/80/index.html on the last KGS bot tournament, I wrote In round 13, Zen9 (as Black) and then oakfoam passed ... . Zen has won: the 14-stone white group ... is dead ... . However they disagreed about the status of the 14-stone group, and the game entered the clean-up phase. Oakfoam again passed, and the game was immediately declared over, with all the stones on the board counted as alive. The game was therefore counted as a win for oakfoam. It seems to me that Black was never given a chance to show that it could capture the dead group. I shall check, and if my understanding is correct, I shall report this to KGS programmer 'wms' as a defect in the clean-up mechanism. I would be grateful if someone could check my interpretation of what happened. Is this just an example of one of the bots handling the clean-up wrong? Or is it the server itself that got it wrong? If it's the server, I must report it to wms. Nick ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Zen blunder in long mirror go game
(2012/01/17 10:45), Jouni Valkonen wrote: Aja, you are talking something that has zero relevance for this topic. Zen was winning by ½ point, but it miscalculated in late yose and blundered severely. This was serious endgame bug. This thread has absolutely nothing to do with the fact that the game was mirrored, but only that that Zen throw away the game in late yose. I don't think Zen was winning. I have seen in my own games several times, when Zen has blundered in late yose and have lost a won games. I think that it is due that zen is winning by half point or so, but miscalculates and then tries something irrational. It would be good idea to analyze, to see what went wrong with Zen's evaluation of the score. Was that it miscalculated the territory or did it misread the seki thing? In this case, the problem was seki handling. But I already have numerous examples of misevaluation, so it is not a special one. Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Zen blunder in long mirror go game
(2012/01/17 8:43), Michael Williams wrote: If the author looks at it, could you update us? I'm always curious about how such things happen, especially in a bot as strong as Zen. Honestly I don't want to spend a lot of time on this type of problem. -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Questions About Zen
(2012/01/16 8:56), Darren Cook wrote: A couple of questions in the game comments yesterday, about Zen, that seemed to go unanswered: 1. Does it have an opening book? Is it hand-made, or a cache of long MCTS analysis, or something in between? Zen has an opening book for 9x9 only. 2. Does it think on the opponents time? Yes it does. -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] CrazyStone in the 5-dan footsteps of Zen
(2012/01/11 5:19), Jean-loup Gailly wrote: > I know the Zen author occasionally logs in as the bot and censors people who abuse the bot. I do this too for pachi2. But it's a real pain to maintain the censor list. My list currently has 147 accounts (probably representing much fewer distinct persons). I did that for only one account, who was an obvious sandbagger. -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] replacing dynamic komi with a scoring function
(2012/01/09 9:56), Jouni Valkonen wrote: There is indeed a problem with Dynamic komi with Zen. Zen often loses the handicap games if black tries to minimize the move count. Often if it is possible to bring game to small yose in around move 180 or so and if not too much behind, then Zen most likely will lose. I have played few games where Zen noted only when filling last dames that it is losing the game by few or half points and then resign. Although, one game was that I lost by ½ points, because I accidentally defended unnecessarily instead of taking the last dame. One game was that i was about ten points behind around move 180, but then Zen played a slack small yose, and lost by 2½ points. Also good and very easy strategy against zen in handicap games is to take all the sides and give center territory to the Zen. Zen almost always will take the center territory as too small and gives sides as too big. I think people often confuse the evaluation problem and the dynamic komi problem. A more urgent problem is the underestimation of the edge and corner territory. -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Zen19D is blitz 5d in KGS
Zen has some code to handle very simple semeais in the playouts. But I am not satisfied with the code. To solve the semeai issue fully, we will need a better algorithm to share the tree information with the simulation. -- Yamato (2011/06/02 3:42), David Fotland wrote: It looks like Zen has solved the issue with semeai in the playouts, but I have no idea how. David -Original Message- From: computer-go-boun...@dvandva.org [mailto:computer-go- boun...@dvandva.org] On Behalf Of Petr Baudis Sent: Wednesday, June 01, 2011 6:29 AM To: Aja; computer-go@dvandva.org Subject: Re: [Computer-go] Zen19D is blitz 5d in KGS Hi! On Wed, Jun 01, 2011 at 11:52:07AM +0800, Aja wrote: Zen19D, running on a small cluster with 26 cores, is rated 5d in KGS now. I played several games with it and felt Zen is getting close to my level. I am happy to see that Crazy Stone and Zen are both so strong. Looks like in computer Go community we got 1-2 stones improvement in the past year. Yes, I just wish people would also publish their improvements. ;-) My guess is that the main improvement of Zen and CrazyStone is in better domain-specific tactical heuristics? Petr "Pasky" Baudis ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] ManyFaces vs Aya today (round 8 of the slow bottournament)
(2011/05/25 22:44), Stefan Kaitschick wrote: As to the problem of smarter playouts, has something along the lines of the "killer heuristic"(the most successful response sofar to a specific move) , used in chess programming, been tried? There's a lot more housekeeping than with RAVE and AMAF, but I can't believe it hasn't been tried. Maybe the effort could be reduced by only storing the response-successrates of the 5*5 surrounding area, or something like that. Anyways, I haven't even heard of a failed attempt, which is a little strange to me. I suppose that is called "Adaptive Playout". Hendrik Baier reported LGRF heuristics and other lots of failed methods. www.ke.tu-darmstadt.de/lehre/arbeiten/master/2010/Baier_Hendrik.pdf -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Fwd: News on Tromp-Cook ?
Aja wrote: >> I suppose that probabilistic simulation on the whole board is not suitable >> for some sort of heuristics, like seki, long semeai or ladder-safe. >> Does Erica handle those in the playouts correctly? > >It's not true. Erica handle seki, semeai and ladder in the playouts combined >with probabilistic simulation. Everything Mogo-type, fixed-sequence >simulation can do, should also be able to be done in probabilistic >simulation. Probabilistic simulation has more flexibility, as Olivier said. That means, you track seki or semeai status incrementally? Or, do you use a resampling algorithm that avoids to play seki-dame or something useless? -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Fwd: News on Tromp-Cook ?
Aja wrote: >> Zen uses sequence-like AND probabilistic simulation. >> Basically it plays around the previous move randomly like MoGo, and these >> moves are biased by gamma values like Crazy Stone. >> >> I am also trying to use probabilistic simulation on the whole board, but >> it does not yet succeed. The main problem is how to combine the semeai, >> seki and useless-move detection with it. > > Very surprised that Zen is using partly sequence-like simulation like >Mogo. Erica uses probabilistic simulation completely. I am not sure if my >implementation is good enough to be called "succeed" , but I am sure that >you will succeed. I suppose that probabilistic simulation on the whole board is not suitable for some sort of heuristics, like seki, long semeai or ladder-safe. Does Erica handle those in the playouts correctly? -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Fwd: News on Tromp-Cook ?
Aja wrote: >> Yamato and I have independently started development of Zen and FudoGo >> (resp) with sharing the same idea*, combining MoGo's sequence-like >> simulation policy and CrazyStone's larger patterns, though later Zen >> uses more complicated policy. >> >> *This is one of the reasons I have started the joint project "DeepZen" >> with Yamato. > > If Zen is not using probabilisic simulation (even with a lot of rules to >make it less-random, like I do in Erica), I will be very surprised with its >performance in few simulations. But anyway, I feel sorry to say anything >about Zen that I am not really sure. Zen uses sequence-like AND probabilistic simulation. Basically it plays around the previous move randomly like MoGo, and these moves are biased by gamma values like Crazy Stone. I am also trying to use probabilistic simulation on the whole board, but it does not yet succeed. The main problem is how to combine the semeai, seki and useless-move detection with it. -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Exploration formulas for UCT
Aja wrote: >> We use the Silver formula: >> >> rave_visits / (rave_visits + real_visits + rave_visits * real_visits * >> 3000) >> >> The figure of 3000 is surprisingly resilient. Even with radically >> different heuristics and playouts, it stays the empirical optimum. > > Interesting. According to Sylvain's original post here, that means you >set bias to sqrt(3000/4)=27.386... But is not bias should be in the range >[0,1]? I guess it should be not "* 3000" but "/ 3000". Zen also uses this type of formula, but the constant value is rather small. I use 400 for the latest version of Zen. -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] UEC2010 tournament pairing
Aja wrote: >> >Fuego won the tournament. >> >> Congratulations to Fuego. >> Was it the same version as in SVN or do you have any secret improvements? >> What was the command line you used? > >Hi Yamato, > >I asked Professor Mueller and Rich the same question yesterday. They replied >that this version of Fuego is almost identical to the one in Kanazawa and >the source will be released soon. My feeling is Fuego's opening play is >obviously much stronger than in Kanazawa. Well, it was almost same as the one in Kanazawa, and, was the Kanazawa version almost same as SVN version or not? >Fuego is scaled up to 112 threads quite well. Rich's paper of CG2010 "On the >Scalability of Parallel UCT" describes the details of how it works. Thanks for the info. -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] UEC2010 tournament pairing
>Fuego won the tournament. Congratulations to Fuego. Was it the same version as in SVN or do you have any secret improvements? What was the command line you used? -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] OS and scoring
Pete Schwamb wrote: >Also, my server currently only support japanese scoring (via gnugo). Is it >acceptable to have >gnugo be the judge of the final score, and would your bot support territory >scoring? Could we do something if gnugo's judgement was wrong? -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Exhibition match
>> Erica is playing a 12 yr old 1 dan professional, >> granddaughter of Fujisawa Shuko, with 6 stone handicap. > >Only right now I finished my sleeping session :o) >So I missed the exhibition match. > >What was the outcome? >Can someone provide the sgf, please? Fujisawa 1p won. The recorded video is available. http://www.ustream.tv/recorded/9944792 -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Monte-Carlo Simulation Balancing in Practice
Aja wrote: >At that time, Erica's strength on 19x19 was still weak, as showned in the >paper. So, I let Erica play against GnuGo 3.8 Level 0, not Level 10. Also, >Level 0 plays much faster. Another important reason is, we were trying to >make the winning rate closer to 50% to get better statistical observation. > >I got very much improvement on 19x19 for past several months. Is the improvement on 19x19 related to Simulation Balancing? -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Last night: Zen5.8 vs John Tromp
I attach the image of Zen's evaluation in the game. Look at the white groups at the top and top-left. Both of them are not 100% correctly evaluated. When this happens, Zen gets weaker because of the divergence and inaccuracy of Monte-Carlo simulations. I think it must be the current major problem of all MC programs. Christian Nentwich wrote: >From a quick look, I would say the bots' evaluation in this case is entirely >correct. Zen was quite far ahead, and then made some bad endgame mistakes. >The trade around move 200 is particularly costly. Playing even a simple safe >move at 176 (like B10) would win the game without trouble. > >Christian > > >On 26 August 2010 10:01, Darren Cook wrote: > >> > Tonight an interesting game was played in the >> > cellars of KGS: Zen19N vs Tromp. >> > (Zen is a bot, Tromp is John Tromp...) >> > >> > You can download the sgf from the KGS archives at >> > http://www.gokgs.com/gameArchives.jsp?user=tromp&year=2010&month=8 >> > >> > For a long time bots (Zen itself, Many Faces) believed >> > that White (=Zen) was clearly ahead. But around move 200 >> > the evaluations swung over, until Tromp won after move 271. >> >> Thanks Ingo. What are people's thoughts: were the programs mistaken in >> their optimism, or did Zen make a mistake in the middle game? >> >> Darren >> >> >> -- >> Darren Cook, Software Researcher/Developer >> >> http://dcook.org/gobet/ (Shodan Go Bet - who will win?) >> http://dcook.org/work/ (About me and my work) >> http://dcook.org/blogs.html (My blogs and articles) >> ___ >> Computer-go mailing list >> Computer-go@dvandva.org >> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >> > inline file >___ >Computer-go mailing list >Computer-go@dvandva.org >http://dvandva.org/cgi-bin/mailman/listinfo/computer-go -- Yamato<>___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Many Faces and Japanese rules
Erik van der Werf wrote: >I just checked to be sure; Steenvreter correctly passes in both >examples. I'm speculating, but I think your problem is that you only >adjust for the difference between Japanese and Chinese counting in the >playouts. However, once consecutive passing gets into the tree you >have to correct earlier. > >The way I think of it Japanese rules have two phases: (1) the game >phase and (2) the confirmation phase. In the game phase passing first >can gain a point. In the confirmation phase it can not. The >confirmation phase normally starts after 2 consecutive passes. If the >confirmation phase has not started yet in the tree then the entire >playout can be considered a confirmation phase. Do you have a special algorithm for the confirmation phase in the tree? Could you explain the details? -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Many Faces and Japanese rules
Aja wrote: >Erica plays PASS (win rate=40%) with 20 playouts. I wonder what is Zen's >result? Great. The current version of Zen never plays pass in this situation. -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Many Faces and Japanese rules
Aja wrote: >Maybe Yamato refered to the situation of seki. I believe my approach can >solve all the normal situations. Attached is a simple example that Black's >best move is to pass (the result is B+0.5). > >My approach in such condition works well. The reason is: If Black fills his >own territory, or suicide one stone in White's territory, then all the >playouts will claim White's win if White passes in UCT tree. How about this? (example2.sgf) -- Yamato example2.sgf Description: Binary data ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Many Faces and Japanese rules
Darren Cook wrote: >> I kind of see how to do it in the tree, but can you handle Japanese >> rules even in the playout phase? ... > >Incidentally, I believe Many Faces is the only MC bot capable of doing >this. Even Made In Japan bots such as Zen seem to be using Chinese rules >exclusively in the playouts. Zen also considers the difference of Japanese rules and Chinese rules in the playouts. However it is unstable when the pass is the only winning move. I wonder if MFG can handle it correctly. -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
Re: [Computer-go] Quiescence Search for Go?
Stefan Kaitschick wrote: >Maybe a way to do quiescence in go is to trade moves in result-critical >local fights before doing the actual search. "Local fights" are not entirely local always. It might be better to do something at the ends of the tree before simulations. -- Yamato ___ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
