Re: Re : [computer-go] an idea... computer go program's rank vs time
On Fri, 2007-01-26 at 13:38 +, ivan dubois wrote: However, if you take for example a computer programm that does straight UCT (global UCT, with no sub-areas), then i believe it can not scale well when board size increases. Because the branching would factor increase proportinaly to the size of the board, and therefore the computation time for an equivalent search deapth will increase exponentialy. Any thoughts ? This can be tested directly. In my own experiments 19x19 improves very rapidly in UCT with each doubling of the number of play-outs. Of course someone will say, yes, but that won't continue and I will have no way to refute their intuition. I am presenting real evidence that it scales at least as far as I can test, and nobody has presented any evidence whatsoever to the contrary other than gut feelings. - Don ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: Re : [computer-go] an idea... computer go program's rank vs time
- Original Message From: Don Dailey [EMAIL PROTECTED] This can be tested directly. In my own experiments 19x19 improves very rapidly in UCT with each doubling of the number of play-outs. May I ask the range of number of playouts tested? Have you considered taking up David Doshay's offer and running some multi-computer simulations? With 72 processors whirling away for a month or so, one could have a lot of interesting data. Btw, how are your experiments with D coming along? Is it becoming competitive, performance-wise, with C or C++? It's here! Your new message! Get new email alerts with the free Yahoo! Toolbar. http://tools.search.yahoo.com/toolbar/features/mail/___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: Re : [computer-go] an idea... computer go program's rank vs time
On Fri, 2007-01-26 at 10:22 -0800, terry mcintyre wrote: - Original Message From: Don Dailey [EMAIL PROTECTED] This can be tested directly. In my own experiments 19x19 improves very rapidly in UCT with each doubling of the number of play-outs. May I ask the range of number of playouts tested? Have you considered taking up David Doshay's offer and running some multi-computer simulations? With 72 processors whirling away for a month or so, one could have a lot of interesting data. Yes, David agreed to let me use his system several weeks ago but I have procrastinated due to many other projects and my work schedule not to mention the time I have wasted posting on this group. Part of my procrastination is that I'm not sure how to make UCT scale to a large number of CPU's.I am an expert in scaling alpha/beta to a large numbers of processors (I did this with Socrates on 1836 processors a few years ago) but it's different with UCT which is inherently serial. I am also considering doing a public huge UCT scalability study with 19x19 go. My idea is for several programs to set their program up to a fixed level (n play-outs) and set up a CGOS server with really long time controls.But I want to just get a normal 19x19 server running first. Btw, how are your experiments with D coming along? Is it becoming competitive, performance-wise, with C or C++? D is a lovely language but the current compilers generate code that is too slow for my taste.I think this situation will improve in the future and I will keep my eye on D. There is no reason I am aware of that D cannot run as fast as C. - Don __ We won't tell. Get more on shows you hate to love (and love to hate): Yahoo! TV's Guilty Pleasures list. ___ 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 : Re : [computer-go] an idea... computer go program's rank vs time
You missunderstood my point. However, I admit it was not clear. What i wanted to say is this : Given a fixed amount of time, strength of monte-carlo algorithm will decrease exponentialy when boardsize increases. It does not mean that monte-carlo does not scale well with time on 19*19. Of course, it all depends on the referential you use to measure strength. By the way, it is not a critisism towards UCT : actualy I do think monte-carlo is the key to build realy strong programs. I just wanted to state that scaling with computation-time is not the only thing that matters. Scaling with board size matters too. It was not another criticism towards you opinion either. - Message d'origine De : Don Dailey [EMAIL PROTECTED] À : computer-go computer-go@computer-go.org Envoyé le : Vendredi, 26 Janvier 2007, 19h05mn 40s Objet : Re: Re : [computer-go] an idea... computer go program's rank vs time On Fri, 2007-01-26 at 13:38 +, ivan dubois wrote: However, if you take for example a computer programm that does straight UCT (global UCT, with no sub-areas), then i believe it can not scale well when board size increases. Because the branching would factor increase proportinaly to the size of the board, and therefore the computation time for an equivalent search deapth will increase exponentialy. Any thoughts ? This can be tested directly. In my own experiments 19x19 improves very rapidly in UCT with each doubling of the number of play-outs. Of course someone will say, yes, but that won't continue and I will have no way to refute their intuition. I am presenting real evidence that it scales at least as far as I can test, and nobody has presented any evidence whatsoever to the contrary other than gut feelings. - Don ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ ___ Découvrez une nouvelle façon d'obtenir des réponses à toutes vos questions ! Profitez des connaissances, des opinions et des expériences des internautes sur Yahoo! Questions/Réponses http://fr.answers.yahoo.com ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: Re : Re : [computer-go] an idea... computer go program's rank vs time
I see what you are saying. Yes, I agree with you that the strength of these programs will decrease exponentially as board sizes increase. By the way, I haven't been offended by any of these messages and I hope I haven't offended anyone either. This is a very interesting conversation to me and I welcome your thoughts and even if I sound too strong I mean no disrespect to anyone. - Don On Fri, 2007-01-26 at 19:05 +, ivan dubois wrote: You missunderstood my point. However, I admit it was not clear. What i wanted to say is this : Given a fixed amount of time, strength of monte-carlo algorithm will decrease exponentialy when boardsize increases. It does not mean that monte-carlo does not scale well with time on 19*19. Of course, it all depends on the referential you use to measure strength. By the way, it is not a critisism towards UCT : actualy I do think monte-carlo is the key to build realy strong programs. I just wanted to state that scaling with computation-time is not the only thing that matters. Scaling with board size matters too. It was not another criticism towards you opinion either. - Message d'origine De : Don Dailey [EMAIL PROTECTED] À : computer-go computer-go@computer-go.org Envoyé le : Vendredi, 26 Janvier 2007, 19h05mn 40s Objet : Re: Re : [computer-go] an idea... computer go program's rank vs time On Fri, 2007-01-26 at 13:38 +, ivan dubois wrote: However, if you take for example a computer programm that does straight UCT (global UCT, with no sub-areas), then i believe it can not scale well when board size increases. Because the branching would factor increase proportinaly to the size of the board, and therefore the computation time for an equivalent search deapth will increase exponentialy. Any thoughts ? This can be tested directly. In my own experiments 19x19 improves very rapidly in UCT with each doubling of the number of play-outs. Of course someone will say, yes, but that won't continue and I will have no way to refute their intuition. I am presenting real evidence that it scales at least as far as I can test, and nobody has presented any evidence whatsoever to the contrary other than gut feelings. - Don ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ ___ Découvrez une nouvelle façon d'obtenir des réponses à toutes vos questions ! Profitez des connaissances, des opinions et des expériences des internautes sur Yahoo! Questions/Réponses http://fr.answers.yahoo.com ___ 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: Re : [computer-go] an idea... computer go program's rank vs time
- Original Message From: Don Dailey [EMAIL PROTECTED] May I ask the range of number of playouts tested? I'm still curious about this question? Part of my procrastination [ about using 72 processors ] is that I'm not sure how to make UCT scale to a large number of CPU's. I am an expert in scaling alpha/beta to a large numbers of processors (I did this with Socrates on 1836 processors a few years ago) but it's different with UCT which is inherently serial. I surely appreciate the difficulties in adapting algorithms to multiple processors - I may be rusty, but some years ago I worked on Neuralware and multiple transputers, 860s, and so forth. It gets a little hairy! Hasn't Mogo been parallelized to 4 processors? Can this be extended to larger numbers? Due to the problems with heat dissipation at higher clock cycles, we'll probably be working with large numbers of processors per chip in the future, rather than Terahertz uniprocessors. Have a burning question? Go to www.Answers.yahoo.com and get answers from real people who know.___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: Re : [computer-go] an idea... computer go program's rank vs time
I personally would love to see more experimental results and less feelings and intuitions on this list. On 1/26/07, Don Dailey [EMAIL PROTECTED] wrote: On Fri, 2007-01-26 at 11:32 -0800, terry mcintyre wrote: - Original Message From: Don Dailey [EMAIL PROTECTED] May I ask the range of number of playouts tested? I'm still curious about this question? I think I started at 64 play-outs, and kept doubling the number of play-outs to some large number where it took an hour to play a single game. I don't currently have the data, but I am willing to reproduce the experiment. Other MC guys can verify it. I'll set it up on a slow computer I have free and I'll start at 64 simulations on a 19x19 board.I'll play 200 games in pairs, 64 vs 128, 128 vs 256, etc. - Don Part of my procrastination [ about using 72 processors ] is that I'm not sure how to make UCT scale to a large number of CPU's. I am an expert in scaling alpha/beta to a large numbers of processors (I did this with Socrates on 1836 processors a few years ago) but it's different with UCT which is inherently serial. I surely appreciate the difficulties in adapting algorithms to multiple processors - I may be rusty, but some years ago I worked on Neuralware and multiple transputers, 860s, and so forth. It gets a little hairy! Hasn't Mogo been parallelized to 4 processors? Can this be extended to larger numbers? Due to the problems with heat dissipation at higher clock cycles, we'll probably be working with large numbers of processors per chip in the future, rather than Terahertz uniprocessors. __ Any questions? Get answers on any topic at Yahoo! Answers. Try it now. ___ 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: Re : [computer-go] an idea... computer go program's rank vs time
On Fri, 2007-01-26 at 14:43 -0500, Don Dailey wrote: I don't currently have the data, but I am willing to reproduce the experiment. Other MC guys can verify it. I'll set it up on a slow computer I have free and I'll start at 64 simulations on a 19x19 board.I'll play 200 games in pairs, 64 vs 128, 128 vs 256, etc. Ok, I am starting the new study and will report the results after each 200 game match starting at 1k simulations. The way my program works is that level 1 does 1024 simulations, and each new level adds 1024 simulations, so I have to start at 1024 instead of 64. 1024 will be pretty weak but I'll keep building up as I go. Maybe I will get someone to help me with other computers. I have an autotester than manages the games and utilities that report statistics on the results. This is my UCT program, not my older Botnoid series. Two games have already been played and the score is tied, level 1 vs level 2 (1024 sims vs 2048 sims.) - Don ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: Re : [computer-go] an idea... computer go program's rank vs time
On Fri, 2007-01-26 at 14:47 -0500, Chris Fant wrote: I personally would love to see more experimental results and less feelings and intuitions on this list. I agree. I will post my data as I go. Just for reference, this is the the Lazarus program that is currently rated at 1807 on CGOS but running 19x19 games. Current results: Rating Win perc Tot Gms Ave Time Player --- --- -- 1600.075.0004 234.8 0002(1024 playouts) 1400.025.0004 117.5 0001(2048 playouts) Black wins:1 25.0 % White wins:3 75.0 % - Don On 1/26/07, Don Dailey [EMAIL PROTECTED] wrote: On Fri, 2007-01-26 at 11:32 -0800, terry mcintyre wrote: - Original Message From: Don Dailey [EMAIL PROTECTED] May I ask the range of number of playouts tested? I'm still curious about this question? I think I started at 64 play-outs, and kept doubling the number of play-outs to some large number where it took an hour to play a single game. I don't currently have the data, but I am willing to reproduce the experiment. Other MC guys can verify it. I'll set it up on a slow computer I have free and I'll start at 64 simulations on a 19x19 board.I'll play 200 games in pairs, 64 vs 128, 128 vs 256, etc. - Don Part of my procrastination [ about using 72 processors ] is that I'm not sure how to make UCT scale to a large number of CPU's. I am an expert in scaling alpha/beta to a large numbers of processors (I did this with Socrates on 1836 processors a few years ago) but it's different with UCT which is inherently serial. I surely appreciate the difficulties in adapting algorithms to multiple processors - I may be rusty, but some years ago I worked on Neuralware and multiple transputers, 860s, and so forth. It gets a little hairy! Hasn't Mogo been parallelized to 4 processors? Can this be extended to larger numbers? Due to the problems with heat dissipation at higher clock cycles, we'll probably be working with large numbers of processors per chip in the future, rather than Terahertz uniprocessors. __ Any questions? Get answers on any topic at Yahoo! Answers. Try it now. ___ 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/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
