Re: [SPAM] Re: [SPAM] [computer-go] MoGo and Fuego question.
How to activate patterns? Why they are not on on default? They are the default, but they are removed if they are not in the path when running (it can be seen on the stderr - the number of patterns read must be 0). In 19x19 there is a big impact. (In the release, there's no pattern) Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] MoGo and Fuego question.
Hi, Is there a way to set up MoGo so it has the same playing strength independently of the CPU and other factors? For instance fixing number of playouts per move? The same question to Fuego developers. Will I achieve this effect with this command? uct_param_player max_games 10 Thanks, Lukasz ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [SPAM] [computer-go] MoGo Zones
(Sylvain et al. 2006) describes the use of CFG-based zones in random simulations to simulate only the local position and tune the score based on few thousands of simulations of outside of the zone. It doesn't seem the idea is too practical (especially with RAVE, but there seem to be more problems), but I'm wondering if MoGo or anyone is still using it, perhaps in a modified form? Not like that in Mogo. We have some local tool for heavy playouts which use local simulations, but for the moment this version is weaker than the light playout - even with fixed number of simulations. Best regards, Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] MoGo Zones
Hi! (Sylvain et al. 2006) describes the use of CFG-based zones in random simulations to simulate only the local position and tune the score based on few thousands of simulations of outside of the zone. It doesn't seem the idea is too practical (especially with RAVE, but there seem to be more problems), but I'm wondering if MoGo or anyone is still using it, perhaps in a modified form? Thanks, -- Petr Pasky Baudis A lot of people have my books on their bookshelves. That's the problem, they need to read them. -- Don Knuth ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo policy: capture stones anywhere?
On Aug 31, 2009, at 10:12 PM, terry mcintyre wrote: If you maintain a list of strings ( connected groups ) of stones and their liberty counts - or perhaps the actual liberties - it should be fairly quick to find a string with just one liberty. I'm currently using pseudoliberties, so that might be tricky. As David Fotland points out, though, the difference shouldn't be huge. In any case, if I read the explanation correctly, this happens infrequently, if several less-expensive tests fail; the cost would be amortized over many trials. Ah, that helps some. I was testing just this aspect of the policy; if I do the escaping and the patterns first, the hit is not as bad (although worse than I'd like). My current implementation is: - Traverse the board. For each point that is the root of an enemy chain, check if that chain is in atari: - If it has 4 pseudoliberties, it's trivially not in atari. - Otherwise, traverse it looking for liberties, abandoning the search if a second liberty is ever found. - If only one liberty is found, increase the number of stones captured by playing at that liberty by the size of the chain. - From those moves that capture, choose the one that captures the most stones. Is there a better way? Peter Drake http://www.lclark.edu/~drake/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] MoGo policy: capture stones anywhere?
IIRC, Fuego has an urgent rule that captures the last-played stone. That rule applies before any others. Then the mop-up capture rule applies just before playing randomly. Mop-up capture is facilitated by a list of strings in atari. Pebbles keeps such a list, like Many Faces. Having such a list is not as important as it used to be, because the rest of Pebbles has gotten heavier. I strongly recommend reading the Fuego source code. Their ideas and implementations might not be best for Orego, and you may have to express ideas differently because Orego might not have infrastructure that Fuego has. Setting these limitations aside, Fuego provides a reference implementation that has been vetted by world-class AI game programmers and very stronger amateur Go masters, and is known to perform very well. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] MoGo policy: capture stones anywhere?
I use real liberties. I think if you want the playouts to do much computation on groups based on liberty count you should switch to using real liberties. I don't think any of the strong programs use pseudoliberties. From: computer-go-boun...@computer-go.org [mailto:computer-go-boun...@computer-go.org] On Behalf Of Peter Drake Sent: Tuesday, September 01, 2009 7:51 AM To: computer-go Subject: Re: [computer-go] MoGo policy: capture stones anywhere? On Aug 31, 2009, at 10:12 PM, terry mcintyre wrote: If you maintain a list of strings ( connected groups ) of stones and their liberty counts - or perhaps the actual liberties - it should be fairly quick to find a string with just one liberty. I'm currently using pseudoliberties, so that might be tricky. As David Fotland points out, though, the difference shouldn't be huge. In any case, if I read the explanation correctly, this happens infrequently, if several less-expensive tests fail; the cost would be amortized over many trials. Ah, that helps some. I was testing just this aspect of the policy; if I do the escaping and the patterns first, the hit is not as bad (although worse than I'd like). My current implementation is: - Traverse the board. For each point that is the root of an enemy chain, check if that chain is in atari: - If it has 4 pseudoliberties, it's trivially not in atari. - Otherwise, traverse it looking for liberties, abandoning the search if a second liberty is ever found. - If only one liberty is found, increase the number of stones captured by playing at that liberty by the size of the chain. - From those moves that capture, choose the one that captures the most stones. Is there a better way? Peter Drake http://www.lclark.edu/~drake/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] MoGo policy: capture stones anywhere?
But note that Fuego is using a GNU license, so if you incorporate any of the Fuego code into your own app, you will have to make your own app available under the GNU license and distribute source to your customers. David I strongly recommend reading the Fuego source code. Their ideas and implementations might not be best for Orego, and you may have to express ideas differently because Orego might not have infrastructure that Fuego has. Setting these limitations aside, Fuego provides a reference implementation that has been vetted by world-class AI game programmers and very stronger amateur Go masters, and is known to perform very well. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo policy: capture stones anywhere?
On Sep 1, 2009, at 8:11 AM, David Fotland wrote: I don’t think any of the strong programs use pseudoliberties. Interesting! Can those involved with other strong programs verify this? My board code is descended from my Java re-implementation of libEGO. I tried writing one using real liberties earlier, and it was considerably slower in random playouts. Perhaps it's worth the speed hit as playouts get heavier. Incidentally, preliminary experiments indicate that the capture-the- largest-chain heuristic (after escaping and matching patterns) does improve strength, despite the speed cost. I'll run a larger experiment tonight... Peter Drake http://www.lclark.edu/~drake/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo policy: capture stones anywhere?
I never tried pseudoliberties in Valkyria. It actually stores arrays of the liberties in addition to the count. This make programming complex algorithms simple, but perhaps not the most efficient way. -Magnus Quoting Peter Drake dr...@lclark.edu: On Sep 1, 2009, at 8:11 AM, David Fotland wrote: I don?t think any of the strong programs use pseudoliberties. Interesting! Can those involved with other strong programs verify this? My board code is descended from my Java re-implementation of libEGO. I tried writing one using real liberties earlier, and it was considerably slower in random playouts. Perhaps it's worth the speed hit as playouts get heavier. Incidentally, preliminary experiments indicate that the capture-the- largest-chain heuristic (after escaping and matching patterns) does improve strength, despite the speed cost. I'll run a larger experiment tonight... Peter Drake http://www.lclark.edu/~drake/ -- Magnus Persson Berlin, Germany ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo policy: capture stones anywhere?
2009/9/1 Peter Drake dr...@lclark.edu: On Sep 1, 2009, at 8:11 AM, David Fotland wrote: I don’t think any of the strong programs use pseudoliberties. Interesting! Can those involved with other strong programs verify this? My board code is descended from my Java re-implementation of libEGO. I tried writing one using real liberties earlier, and it was considerably slower in random playouts. I started out by looking at Orego's code when I first tried MC ;-). Since then I found that even with very light playouts, pseudo-liberties is only marginally (like a few percent) faster than keeping actual liberty-counts. This performance hit is easily recovered by having the real number of liberties at all times for other parts. Just the coding is a bit more work to make it efficient. But you can check the plug-and-go project for a Java implementation. Mark ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] MoGo policy: capture stones anywhere?
Interesting! Can those involved with other strong programs verify this? Pebbles isn't a particularly strong program, but using real liberty counts is better. I have recently taken Pebbles offline to renovate internals. I am adding lists of liberties, too. I am convinced that the richer representation will improve speed, because I have written ridiculous amounts of code that searches for liberties in the course of trying to analyze semeais, ladders, and so on. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] MoGo policy: capture stones anywhere?
MoGo's playout policy (at one time) is given in section 3.2 of Gelly et al's paper, Modification of UCT with Patterns in Monte-Carlo Go: We describe briefly how the improved random mode generates moves. It first verifies whether the last played move is an Atari; if yes, and if the stones under Atari can be saved (in the sense that it can be saved by capturing stones or increasing liberties), it chooses one saving move randomly; otherwise it looks for interesting moves in the 8 positions around the last played move and plays one randomly if there is any; otherwise it looks for the moves capturing stones on the Go board, plays one if there is any. At last, if still no move is found, it plays one move randomly on the Go board. We've implemented much of this, which has made Orego considerably stronger. The problem is with this part: otherwise it looks for the moves capturing stones on the Go board Does this really mean traversing the entire board looking for captures? Doing so seems to create a catastrophic speed hit. Peter Drake http://www.lclark.edu/~drake/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo policy: capture stones anywhere?
If you maintain a list of strings ( connected groups ) of stones and their liberty counts - or perhaps the actual liberties - it should be fairly quick to find a string with just one liberty. In any case, if I read the explanation correctly, this happens infrequently, if several less-expensive tests fail; the cost would be amortized over many trials. Terry McIntyre terrymcint...@yahoo.com And one sad servitude alike denotes The slave that labours and the slave that votes -- Peter Pindar From: Peter Drake dr...@lclark.edu To: Computer Go computer-go@computer-go.org Sent: Monday, August 31, 2009 9:56:16 PM Subject: [computer-go] MoGo policy: capture stones anywhere? MoGo's playout policy (at one time) is given in section 3.2 of Gelly et al's paper, Modification of UCT with Patterns in Monte-Carlo Go: We describe briefly how the improved random mode generates moves. It first verifies whether the last played move is an Atari; if yes, and if the stones under Atari can be saved (in the sense that it can be saved by capturing stones or increasing liberties), it chooses one saving move randomly; otherwise it looks for interesting moves in the 8 positions around the last played move and plays one randomly if there is any; otherwise it looks for the moves capturing stones on the Go board, plays one if there is any. At last, if still no move is found, it plays one move randomly on the Go board. We've implemented much of this, which has made Orego considerably stronger. The problem is with this part: otherwise it looks for the moves capturing stones on the Go board Does this really mean traversing the entire board looking for captures? Doing so seems to create a catastrophic speed hit. Peter Drake http://www.lclark.edu/~drake/ ___ 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] MoGo policy: capture stones anywhere?
I maintain a list of strings with exactly one liberty. If I disable the code to track one liberty groups and generate capture moves the speed goes up from 17.7 K playouts/s to 18.1K playouts/s, so it's a small difference, probably not statistically significant. David -Original Message- From: computer-go-boun...@computer-go.org [mailto:computer-go- boun...@computer-go.org] On Behalf Of Peter Drake Sent: Monday, August 31, 2009 9:56 PM To: Computer Go Subject: [computer-go] MoGo policy: capture stones anywhere? MoGo's playout policy (at one time) is given in section 3.2 of Gelly et al's paper, Modification of UCT with Patterns in Monte-Carlo Go: We describe briefly how the improved random mode generates moves. It first verifies whether the last played move is an Atari; if yes, and if the stones under Atari can be saved (in the sense that it can be saved by capturing stones or increasing liberties), it chooses one saving move randomly; otherwise it looks for interesting moves in the 8 positions around the last played move and plays one randomly if there is any; otherwise it looks for the moves capturing stones on the Go board, plays one if there is any. At last, if still no move is found, it plays one move randomly on the Go board. We've implemented much of this, which has made Orego considerably stronger. The problem is with this part: otherwise it looks for the moves capturing stones on the Go board Does this really mean traversing the entire board looking for captures? Doing so seems to create a catastrophic speed hit. Peter Drake http://www.lclark.edu/~drake/ ___ 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] MoGo - ManyFaces
13 games were played and the total score was 8-5 for CzechBot. I wonder how would they play if on even grounds. The general game pattern was the usual wild middlegame wrestling typical of MC, with CzechBot usually getting large edge initially (70% winning probability and seemingly unshakeable position), but then in the lost games making a careless blunder or two when it got too much ahead, and panicking subsequently. Overlooking ladders seemed to happen to it several times. trollSo maybe CzechBot should put a komi burden on itself when it gets too optimistic./troll Stefan ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo - ManyFaces
Hi! Today there was a short discussion about the strongest bot currently online on KGS and I got curious whether ManyFaces or CzechBot (bleeding edge MoGo) is stronger, so I made it play against ManyFaces. CzechBot is running as dual-thread pondering MoGo on slightly loaded dual-core Athlon64 4800+ with 4G RAM, so perhaps slightly better hardware than ManyFaces Turns out that ManyFaces is running on a quad-core, so the hardware difference is bigger than I thought and in favor of ManyFaces. Petr Pasky Baudis ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] MoGo - ManyFaces
Hi! Today there was a short discussion about the strongest bot currently online on KGS and I got curious whether ManyFaces or CzechBot (bleeding edge MoGo) is stronger, so I made it play against ManyFaces. CzechBot is running as dual-thread pondering MoGo on slightly loaded dual-core Athlon64 4800+ with 4G RAM, so perhaps slightly better hardware than ManyFaces; it was reading slightly above 3kpps. Since CzechBot is ranked one stone lower than ManyFaces, all the games were played josen: with ManyFaces white taking no komi. The time was 30:00 SD. The ruleset was Japanese and CzechBot was assuming Chinese rules, but in practice it never mattered. 13 games were played and the total score was 8-5 for CzechBot. I wonder how would they play if on even grounds. The general game pattern was the usual wild middlegame wrestling typical of MC, with CzechBot usually getting large edge initially (70% winning probability and seemingly unshakeable position), but then in the lost games making a careless blunder or two when it got too much ahead, and panicking subsequently. Overlooking ladders seemed to happen to it several times. This is just a FYI if anyone is interested in the relative strength or would find the games interesting. -- Petr Pasky Baudis A lot of people have my books on their bookshelves. That's the problem, they need to read them. -- Don Knuth ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] MoGo-4C-Po-100Mnod on CGOS
MoGo-4C-Po-100Mnod does not remove dead stones, and lose. http://cgos.boardspace.net/9x9/SGF/2009/07/20/813050.sgf Please add --playsAgainstHuman 0 option, like this, mogo --9 --nbTotalSimulations 1 --nbThread 1 --pondering 1 --playsAgainstHuman 0 And following bots also don't remove. Fuego4C4PlaPo20Mno(2438) MoGo-4C-Po-100Mnod(2401) GnuGo-mc-10K-lev11(1818) GG-500(1646) Fuego-700node(1508) Fuego-500node(1400) Fuego-300node(1248) Reagrds, Hiroshi Yamashita ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo and passing
On http://www.lri.fr/~gelly/MoGo_Download.htmhttp://www.lri.fr/%7Egelly/MoGo_Download.htm, under the FAQ section, I found the bullet point: MoGo continues playing after the game is over?: MoGo never consider a pass unless you pass first. If you think the game is over, simply pass. Is this still true? If so, how does MoGo deal with situations where the best move is to pass (e.g. seki). When passing is necessary, mogo passes. So this sentence is not exactly true; mogo can pass first in extremal cases (very rare cases, against humans). Best regards, Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo and passing
Hi, Olivier answered for the new version. On the downloadable version, I don't remember exactly (almost 2 years back now...), but I think Mogo will still pass if all the other moves are clearly loosing. So it should understand somehow Seki situations. If that is correct, the sentence is not completely accurate. It should more be: MoGo never consider a pass unless you pass first or all other moves are obviously loosing the game. Sylvain On Wed, Jun 24, 2009 at 10:11 PM, Seth Pellegrinose...@lclark.edu wrote: Hello all, On http://www.lri.fr/~gelly/MoGo_Download.htm , under the FAQ section, I found the bullet point: MoGo continues playing after the game is over?: MoGo never consider a pass unless you pass first. If you think the game is over, simply pass. Is this still true? If so, how does MoGo deal with situations where the best move is to pass (e.g. seki). Thanks, Seth Pellegrino ___ 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] MoGo and passing
Obviously I should read better the emails before answering. Olivier rightly answered for all versions. Sorry, Sylvain On Tue, Jun 30, 2009 at 7:59 PM, Sylvain Gellysylvain.ge...@m4x.org wrote: Hi, Olivier answered for the new version. On the downloadable version, I don't remember exactly (almost 2 years back now...), but I think Mogo will still pass if all the other moves are clearly loosing. So it should understand somehow Seki situations. If that is correct, the sentence is not completely accurate. It should more be: MoGo never consider a pass unless you pass first or all other moves are obviously loosing the game. Sylvain On Wed, Jun 24, 2009 at 10:11 PM, Seth Pellegrinose...@lclark.edu wrote: Hello all, On http://www.lri.fr/~gelly/MoGo_Download.htm , under the FAQ section, I found the bullet point: MoGo continues playing after the game is over?: MoGo never consider a pass unless you pass first. If you think the game is over, simply pass. Is this still true? If so, how does MoGo deal with situations where the best move is to pass (e.g. seki). Thanks, Seth Pellegrino ___ 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] Mogo on supercomputer
When Mogo runs on the supercomputer with long-ish time limits, how big does the search tree get? ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Mogo on supercomputer
When Mogo runs on the supercomputer with long-ish time limits, how big does the search tree get? Plotting the depth/number of nodes as a function of the thinking time might be a good idea... No idea :-( I just remember that changing the number of visits before adding a new node in the tree changes the length of the simulated variations without having a strong impact on the level. Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] Mogo MCTS is not UCT ? (GENERALIZED_AMAF)
In the mentioned article it is said that : A weakness remains, due to the initial time: at the very beginning, neither AMAF values nor standard statistics are available I have developed a model that aim at healing that part. It gives virtual_simulations out of arbitrary ancestor at a rate of 1/2 for each node you go through. I called that process GENERALIZED_AMAF. I'm not familiar with mathematical terms, but i think the idea is dead simple : AMAF goes as this : Evaluate the expected results KNOWING that move A has been played. GENERALIZED_AMAF : Evaluate the expected results KNOWING that move A (AND) move B (AND) move C has been played. -- You can easily build for example an AMAF_map, which would get both the moves that have been marked as played by black on one part, and the moves that has been played by white in the other part. That is associated with the result (black win/lose) Given a node, classical AMAF would then try to check out every of those maps, for each simulation starting from that node where the child you are trying to score has been played. Generalized AMAF can build statistics out simulation from arbitrary ancestors from the considered node. You would get (1/2)^n simulations from the ancestor as GENERALIZED_AMAF_simulations (n the number of choice made from the ancestor to the node you assess). Suppose that you have 1000 simulations for a root node R. Then amaf gives you about 500 simulations for every child let's call them Cn those child, where n is the id of the child. Cn also represent the move made from R to reach that child. Then for each child of Cn, you can get 250 generalized_amaf simulations. you would consider from the set of simulations from the root, only those where Cn has been played, and then aggregate the results in the AMAF fashion for each son of Cn. My prototype was scoring 30% win against Gnu-go lvl 0, without any tuning using plain standard_light_simulations. It would use the generalized-amaf as a way to get RAVE values. Then it would guarantees that the most promising nodes is explored exponentially more. (it used a raw non deterministic algorithm) I did not however tried to compare that win ratio, with the win ratio i would have got out of standard Amaf. Best regard, Denis FIDAALI. Date: Mon, 1 Dec 2008 21:55:03 +0100 From: [EMAIL PROTECTED] To: computer-go@computer-go.org Subject: Re: [computer-go] Mogo MCTS is not UCT ? I think it's now well known that Mogo doesn't use UCT. I realize that i have no idea at all what Mogo do use for it's MCTS. A complicated formula mixing (i) patterns (ii) rules (iii) rave values (iv) online statistics Also we have a little learning (i.e. late parts of simulations are evolved based on online statistics and not only the early parts). I really wonder if there was an article describing the new MCTS of mogo somewhere that i missed. How is it better than UCT ? http://www.lri.fr/~teytaud/eg.pdf contains most of the information (many other things have been tried and kept as they provided small improvements, but essentially the ideas are in this version) Best regards, Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ _ Téléphonez gratuitement à tous vos proches avec Windows Live Messenger ! Téléchargez-le maintenant ! http://www.windowslive.fr/messenger/1.asp___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Mogo MCTS is not UCT ? (GENERALIZED_AMAF)
At first blush, it sounds like a reasonable idea. However, as always with these things, the proof of the cake is in the eating. So you'd have to try it out against a ref-bot to see if it improves things. You also need to keep an eye on time used, as it sounds a lot more CPU intensive than plain AMAF. As to the quote below: what is meant with the 'very beginning'? If that's the very beginning of the game, then I suppose an opening book can be used for the initial statistics. If it means, at the beginning of the search process of a move, then I suppose you can start by using the data generated by the previous search. Most likely it's very rare your opponent plays a move that has not been investigated by your previous search at all. Another thing you can do is use the opponent's time to do a little preparation. I haven't spent time looking at 'pondering' yet, but if it were me I'd start by building a two-ply tree of AMAF values. Say you expand the tree after N playouts. You do N simulations at the current level. Then you choose the best node and do N simulations there. Then at the second-best node, etc. After N*m (m=number of legal moves) simulations, you have the initial AMAF data for every possible move your opponent can play. Even with N fairly large you're likely to be able to finish this process before your opponents plays his move. Of course the time saved is very little, as will always be the case with pondering. The 'weakness' as seen in the article is a very small one. Mark On 2-dec-08, at 06:48, Denis fidaali wrote: In the mentioned article it is said that : A weakness remains, due to the initial time: at the very beginning, neither AMAF values nor standard statistics are available I have developed a model that aim at healing that part. It gives virtual_simulations out of arbitrary ancestor at a rate of 1/2 for each node you go through. I called that process GENERALIZED_AMAF. I'm not familiar with mathematical terms, but i think the idea is dead simple : AMAF goes as this : Evaluate the expected results KNOWING that move A has been played. GENERALIZED_AMAF : Evaluate the expected results KNOWING that move A (AND) move B (AND) move C has been played. -- You can easily build for example an AMAF_map, which would get both the moves that have been marked as played by black on one part, and the moves that has been played by white in the other part. That is associated with the result (black win/lose) Given a node, classical AMAF would then try to check out every of those maps, for each simulation starting from that node where the child you are trying to score has been played. Generalized AMAF can build statistics out simulation from arbitrary ancestors from the considered node. You would get (1/2)^n simulations from the ancestor as GENERALIZED_AMAF_simulations (n the number of choice made from the ancestor to the node you assess). Suppose that you have 1000 simulations for a root node R. Then amaf gives you about 500 simulations for every child let's call them Cn those child, where n is the id of the child. Cn also represent the move made from R to reach that child. Then for each child of Cn, you can get 250 generalized_amaf simulations. you would consider from the set of simulations from the root, only those where Cn has been played, and then aggregate the results in the AMAF fashion for each son of Cn. My prototype was scoring 30% win against Gnu-go lvl 0, without any tuning using plain standard_light_simulations. It would use the generalized-amaf as a way to get RAVE values. Then it would guarantees that the most promising nodes is explored exponentially more. (it used a raw non deterministic algorithm) I did not however tried to compare that win ratio, with the win ratio i would have got out of standard Amaf. Best regard, Denis FIDAALI. Date: Mon, 1 Dec 2008 21:55:03 +0100 From: [EMAIL PROTECTED] To: computer-go@computer-go.org Subject: Re: [computer-go] Mogo MCTS is not UCT ? I think it's now well known that Mogo doesn't use UCT. I realize that i have no idea at all what Mogo do use for it's MCTS. A complicated formula mixing (i) patterns (ii) rules (iii) rave values (iv) online statistics Also we have a little learning (i.e. late parts of simulations are evolved based on online statistics and not only the early parts). I really wonder if there was an article describing the new MCTS of mogo somewhere that i missed. How is it better than UCT ? http://www.lri.fr/~teytaud/eg.pdf contains most of the information (many other things have been tried and kept as they provided small improvements, but essentially the ideas are in this version) Best regards, Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ Votre correspondant a choisi Hotmail et
[computer-go] Mogo MCTS is not UCT ?
I think it's now well known that Mogo doesn't use UCT. I realize that i have no idea at all what Mogo do use for it's MCTS. There are only two things i dislike about UCT : - It's slow to compute. - It's deterministic I really wonder if there was an article describing the new MCTS of mogo somewhere that i missed. How is it better than UCT ? _ Email envoyé avec Windows Live Hotmail. Dites adieux aux spam et virus, passez à Hotmail ! C'est gratuit ! http://www.windowslive.fr/hotmail/default.asp___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Mogo MCTS is not UCT ?
On Dec 1, 2008, at 3:38 AM, Denis fidaali [EMAIL PROTECTED] wrote: I think it's now well known that Mogo doesn't use UCT. I realize that i have no idea at all what Mogo do use for it's MCTS. There are only two things i dislike about UCT : - It's slow to compute. - It's deterministic I really wonder if there was an article describing the new MCTS of mogo somewhere that i missed. How is it better than UCT ? My understanding is that MoGo dropped the upper confidence bound portion. That makes it a bit faster, but still deterministic for a given set of playout results. Heuristics and RAVE give a sufficiently good move ordering that less exploration is needed. IIRC, Valkyra still uses UCT, but has a very low coefficent on the upper confidence bound term. Qui vous permet d'enregistrer la TV sur votre PC et lire vos emails sur votre mobile ? la réponse en vidéo la réponse en vidéo ___ 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] Mogo MCTS is not UCT ?
Let's assume that the UCT formula is UCTValue(parent, n) = winrate + sqrt((ln(parent.visits))/(5*n.nodevisits)) (taken from sensei library) What is the Upper confidence bound term ? That would'nt be sqrt((ln(parent.visits))/(5*n.nodevisits)) ?? I doubt that exploring only the move with the best winrate would lead to a fast enough convergence even on 9x9. Is that what you meant by dropping the upper confidence bound term ? Otherwise, what does the formula without the upper confidence bound term do looks like ? My understanding is that MoGo dropped the upper confidence bound portion. That makes it a bit faster, but still deterministic for a given set of playout results. Heuristics and RAVE give a sufficiently good move ordering that less exploration is needed. IIRC, Valkyra still uses UCT, but has a very low coefficent on the upper confidence bound term. _ Email envoyé avec Windows Live Hotmail. Dites adieux aux spam et virus, passez à Hotmail ! C'est gratuit ! http://www.windowslive.fr/hotmail/default.asp___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Mogo MCTS is not UCT ?
I think it's now well known that Mogo doesn't use UCT. I realize that i have no idea at all what Mogo do use for it's MCTS. A complicated formula mixing (i) patterns (ii) rules (iii) rave values (iv) online statistics Also we have a little learning (i.e. late parts of simulations are evolved based on online statistics and not only the early parts). I really wonder if there was an article describing the new MCTS of mogo somewhere that i missed. How is it better than UCT ? http://www.lri.fr/~teytaud/eg.pdf contains most of the information (many other things have been tried and kept as they provided small improvements, but essentially the ideas are in this version) Best regards, Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Mogo MCTS is not UCT ?
On 1-dec-08, at 18:55, Olivier Teytaud wrote: I think it's now well known that Mogo doesn't use UCT. I realize that i have no idea at all what Mogo do use for it's MCTS. A complicated formula mixing (i) patterns (ii) rules (iii) rave values (iv) online statistics Isn't that technically still UCT? I mean, you use different input and probably a different formula, but most likely what you do is still establish an upper bound to which extent you trust the win-ratio (and possibly other data) to determine which node to extend next. When that upper-bound is passed you decide to extend a less promising node to make sure you don't overlook an unlikely but possibly very good candidate. It's just that people here have come to associate UCT with a particular formula, but that formula is not the only way you can establish an upper confidence bound. Mark ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] Mogo Opening, Building Strategy ?
Hi. Is there any theoretical reasons for the Mogo Opening being built out of self play, rather than by spending time increasing the number of simulations at the root, and after a time, keeping what seems to be the best ? Obviously one could object that increasing the number of simulations would lead to out of memory. But then it seems that there are ways of killing the less explored branches. Another thing also, is that it may be far easier to get a massive parallel processing of self play. So i really wondered what would be best, both for time efficiency, and strength : spending a lot of computer power refining the tree, or using self play from a position, to assess its value ? What was the reason behind the choice of the Mogo team ? _ Inédit ! Des Emoticônes Déjantées! Installez les dans votre Messenger ! http://www.ilovemessenger.fr/Emoticones/EmoticonesDejantees.aspx___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Mogo Opening, Building Strategy ?
Is there any theoretical reasons for the Mogo Opening being built out of self play, rather than by spending time increasing the number of simulations at the root, and after a time, keeping what seems to be the best ? There are practical reasons: our approach can be used with humans or other programs as opponent as well; we can benefit from games launched for other purposes than opening-building; and we can easily parallelize the algorithm on grids. No other reason, at least for me - but these reasons are enough I guess. The alternate approach is nice, but is difficult to use for tenths of years of CPU - whereas using preemptable mode in grids, we can have access to a huge computational power. From a more technical point of view, I think that the idea of using results of games of strong versions of mogo is better for avoiding biases in the MC. But it's only a conjecture. Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] Mogo Opening, Building Strategy ?
Thanks a lot for your quick answer. By conjecture, i suppose you mean that no experiments yet has been ran as to assess this hypothesis ? I think Sylvain (and maybe just everyone else) has tried at some point to use a UCT decision bot, as a way to get the simulation done. Then using those high level simulations in an other UCT decision tree (or AMAF, or FirstMove wining stats) From what i recall, the results were disappointing. Also don dailey tried to build an AMAF over a bot that would use AMAF as a decision with very little expectation that this would lead to anything worthy. I don't know how hard Sylvain tried at his time. Yet you have this feeling that using High level mogo games as a way to get a simulation done could lead to interesting results. I also have this feeling. For example, it is well known that Mogo-style decision (or crazy stone) lead to very poor understanding of seki (and/or semeai ?) Would'nt the use of high level game as simulation get to better understanding of those really nasty situations ? Then, i guess that if nobody has ever run any experiments, as to get measure of the efficiency of increasing the UCT tree against using high-level-simulation, there must be a reason ... Is it that that it is known it would consume to much time and resources ? Is it that knowing the results of this measure would prove of little value ? If there is a point where high-level-simulations really give a stronger evaluation function, wouldn't it be good to know about it ? Date: Sun, 30 Nov 2008 10:10:14 +0100 From: [EMAIL PROTECTED] To: computer-go@computer-go.org Subject: Re: [computer-go] Mogo Opening, Building Strategy ? Is there any theoretical reasons for the Mogo Opening being built out of self play, rather than by spending time increasing the number of simulations at the root, and after a time, keeping what seems to be the best ? There are practical reasons: our approach can be used with humans or other programs as opponent as well; we can benefit from games launched for other purposes than opening-building; and we can easily parallelize the algorithm on grids. No other reason, at least for me - but these reasons are enough I guess. The alternate approach is nice, but is difficult to use for tenths of years of CPU - whereas using preemptable mode in grids, we can have access to a huge computational power. From a more technical point of view, I think that the idea of using results of games of strong versions of mogo is better for avoiding biases in the MC. But it's only a conjecture. Olivier _ Email envoyé avec Windows Live Hotmail. Dites adieux aux spam et virus, passez à Hotmail ! C'est gratuit ! http://www.windowslive.fr/hotmail/default.asp___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Mogo Opening, Building Strategy ?
By conjecture, i suppose you mean that no experiments yet has been ran as to assess this hypothesis ? Yes. The other reasons were sufficient :-) I think Sylvain (and maybe just everyone else) has tried at some point to use a UCT decision bot, as a way to get the simulation done. Then using those high level simulations in an other UCT decision tree (or AMAF, or FirstMove wining stats) From what i recall, the results were disappointing. At least, it has been clearly established that replacing the random player by a stronger player does not imply the Monte-Carlo program built on top of the random player becomes stronger (even with fixed number of simulations) But, it is also clearly established that the building of the opening book by self-play clearly works, whereas it is roughly the same idea. I guess the reason is the difference of strength of the player - a MCTS (Monte-Carlo Tree Search - I don't write UCT here as it is not UCT in mogo) built on top of a perfect player should be a perfect player (this is formally obvious). So perhaps for huge computational power, this approach (building MCTS on top of MCTS) is consistent. For example, it is well known that Mogo-style decision (or crazy stone) lead to very poor understanding of seki (and/or semeai ?) Would'nt the use of high level game as simulation get to better understanding of those really nasty situations ? I hope so, at least for 9x9 and with really huge computational power. (by the way I'm afraid we have to patch semeai manually) Is it that that it is known it would consume to much time and resources ? I think it is really difficult to do that - you have to dump your results unless you have a machine for a huge time without any reboot, also if you want to have a huge computational effort you have to parallelize it - I think this is really hard. But perhaps trying mogo against mogo built on top of a mogo with e.g. 1s/move would be fine... this is easy to organize. Well, it requires some time, and time is always expensive :-) Best regards, Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Mogo Opening, Building Strategy ?
On Sun, 2008-11-30 at 13:38 +0100, Olivier Teytaud wrote: But, it is also clearly established that the building of the opening book by self-play clearly works, whereas it is roughly the same idea. I guess the reason is the difference of strength of the player - a MCTS (Monte-Carlo Tree Search - I don't write UCT here as it is not UCT in mogo) built on top of a perfect player should be a perfect player (this is formally obvious). So perhaps for huge computational power, this approach (building MCTS on top of MCTS) is consistent. I've always had this idea that the best way to build an opening book is the best way to build a general playing engine. You are trying to solve the same exact problem - what is the best move in this position? - Don signature.asc Description: This is a digitally signed message part ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Mogo Opening, Building Strategy ?
From: Don Dailey [EMAIL PROTECTED] I've always had this idea that the best way to build an opening book is the best way to build a general playing engine. You are trying to solve the same exact problem - what is the best move in this position? When building an opening book, you have the advantage of not playing against the clock. In fact, a good opening book ( one which your game-playing engine knows how to use ) can shave time during the game itself. Knows how to use can be subtle; many joseki depend on knowing the exact line of play, the move choices depend on knowing the exact ( not approximate ) results of ladders and semeai. Against better players, approximate answers tend toward disaster. A book move might work sometimes, not others, and the program won't know the difference. I think the opening book and general playing engine solve similar problems: what is the best move which can be discovered with finite resources? The opening problem must solve an additional side constraint: it must suggest moves which can be correctly exploited by the playing engine, which may have less time and computational power available. A sufficiently broad and deep book can make up for lack of computational resources during the game; such a book needs to know the best refutations for each of many possible plays. Joseki and fuseki books for humans are full of annotations: move a is used when the ladder works; otherwise, b is recommended; c is a mistake, which can be exploited by ... A book for computer programs might need similar annotations. Some programs may need different books, depending on whether they have a fast or slow clock, one or many processors. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Mogo Opening, Building Strategy ?
It's true that building an opening book in an automated way can be done off-line which gives us more resources. That's really the basis for this thought that we are trying to solve the same problem. As a thought experiment, imagine some day in the future, when computers are 1 thousand times faster. What would you do with that extra time to make your program more scalable?You might be able to use your book building algorithm, whatever it might be, to find moves for you. If there is a better way, then you would use it instead. But if there is a better way why are you not using it for creating book moves? Of course the details of how to do it may not be exactly the same - because we may be forced to simulate in memory hash tables with disk space for instance, or to deal with other real-world constraints. However I think that at least in principle, we should be doing the same thing. MCTS really feels to me like a superb book building algorithm. Computer Chess books (at least the automated part) are built essentially by taking millions of games from master play and picking out the ones that seem to work best. Those games are like playouts. The moves that score the best are played the most. We have a kind of MCTS here. - Don On Sun, 2008-11-30 at 09:15 -0800, terry mcintyre wrote: From: Don Dailey [EMAIL PROTECTED] I've always had this idea that the best way to build an opening book is the best way to build a general playing engine. You are trying to solve the same exact problem - what is the best move in this position? When building an opening book, you have the advantage of not playing against the clock. In fact, a good opening book ( one which your game-playing engine knows how to use ) can shave time during the game itself. Knows how to use can be subtle; many joseki depend on knowing the exact line of play, the move choices depend on knowing the exact ( not approximate ) results of ladders and semeai. Against better players, approximate answers tend toward disaster. A book move might work sometimes, not others, and the program won't know the difference. I think the opening book and general playing engine solve similar problems: what is the best move which can be discovered with finite resources? The opening problem must solve an additional side constraint: it must suggest moves which can be correctly exploited by the playing engine, which may have less time and computational power available. A sufficiently broad and deep book can make up for lack of computational resources during the game; such a book needs to know the best refutations for each of many possible plays. Joseki and fuseki books for humans are full of annotations: move a is used when the ladder works; otherwise, b is recommended; c is a mistake, which can be exploited by ... A book for computer programs might need similar annotations. Some programs may need different books, depending on whether they have a fast or slow clock, one or many processors. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ signature.asc Description: This is a digitally signed message part ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Mogo Opening, Building Strategy ?
- Original Message From: Don Dailey [EMAIL PROTECTED] MCTS really feels to me like a superb book building algorithm. Computer Chess books (at least the automated part) are built essentially by taking millions of games from master play and picking out the ones that seem to work best. Those games are like playouts. The moves that score the best are played the most. We have a kind of MCTS here. Interesting, the book moves are not generated by random playouts, but by professional ( or highly skilled ) players. In the Chess world, what is meant by picking out the ones that seem to work best? My impression is that computer go programs do not, at this stage in their evolution, make good use of professional book moves; too much professional knowledge is actually in the part of the tree which is almost never played in pro-pro games - the how to beat up on mistakes and what not to do when playing against a pro parts of the tree. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Mogo Opening, Building Strategy ?
On Sun, 2008-11-30 at 14:33 -0800, terry mcintyre wrote: - Original Message From: Don Dailey [EMAIL PROTECTED] MCTS really feels to me like a superb book building algorithm. Computer Chess books (at least the automated part) are built essentially by taking millions of games from master play and picking out the ones that seem to work best. Those games are like playouts. The moves that score the best are played the most. We have a kind of MCTS here. Interesting, the book moves are not generated by random playouts, but by professional ( or highly skilled ) players. Many chess opening books are created by a statistical analysis of lots of human games. Some books are created by hand. Very tediously. Even the ones created with the aid of human games are sometimes modified by hand, at least the top notch books. But in the context of this discussion we note that books can and are created solely from databases of top quality games. You can get a reasonable book that way. In the Chess world, what is meant by picking out the ones that seem to work best? What I mean is that you look at the statistics of the moves and base your opening book on the moves that gave the best results. You can also go by the moves that are played the most - with the assumption that if they are played a lot they must be good. It is typical to do a combination of both - if it's played a lot and also scores good, use it. I think some have tried mini-max too. It's possible that a move seems to has great success in general, but not if it's responded to in a certain way. It could be that in recent months or years a refutation has been found, and that a move that used to work really well has been found to be bad. My impression is that computer go programs do not, at this stage in their evolution, make good use of professional book moves; too much professional knowledge is actually in the part of the tree which is almost never played in pro-pro games - the how to beat up on mistakes and what not to do when playing against a pro parts of the tree. Even in chess, despite the awesome strength of the programs, human knowledge of the openings still reigns supreme, although it's now the case that computers are helping to build opening theory by finding new moves - in chess these are called theoretical novelties and computers have produced many of them from what I understand. - Don ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ signature.asc Description: This is a digitally signed message part ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Mogo Opening, Building Strategy ?
You can read about some such novelties found using Rybka here: http://www.rybkachess.com/index.php?auswahl=Rybka+3+book Don Dailey wrote: On Sun, 2008-11-30 at 14:33 -0800, terry mcintyre wrote: - Original Message From: Don Dailey [EMAIL PROTECTED] MCTS really feels to me like a superb book building algorithm. Computer Chess books (at least the automated part) are built essentially by taking millions of games from master play and picking out the ones that seem to work best. Those games are like playouts. The moves that score the best are played the most. We have a kind of MCTS here. Interesting, the book moves are not generated by random playouts, but by professional ( or highly skilled ) players. Many chess opening books are created by a statistical analysis of lots of human games. Some books are created by hand. Very tediously. Even the ones created with the aid of human games are sometimes modified by hand, at least the top notch books. But in the context of this discussion we note that books can and are created solely from databases of top quality games. You can get a reasonable book that way. In the Chess world, what is meant by picking out the ones that seem to work best? What I mean is that you look at the statistics of the moves and base your opening book on the moves that gave the best results. You can also go by the moves that are played the most - with the assumption that if they are played a lot they must be good. It is typical to do a combination of both - if it's played a lot and also scores good, use it. I think some have tried mini-max too. It's possible that a move seems to has great success in general, but not if it's responded to in a certain way. It could be that in recent months or years a refutation has been found, and that a move that used to work really well has been found to be bad. My impression is that computer go programs do not, at this stage in their evolution, make good use of professional book moves; too much professional knowledge is actually in the part of the tree which is almost never played in pro-pro games - the how to beat up on mistakes and what not to do when playing against a pro parts of the tree. Even in chess, despite the awesome strength of the programs, human knowledge of the openings still reigns supreme, although it's now the case that computers are helping to build opening theory by finding new moves - in chess these are called theoretical novelties and computers have produced many of them from what I understand. - 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] MoGo v.s. Kim rematch (Jason House's paper)
Sent from my iPhone On Sep 24, 2008, at 5:16 PM, Jason House [EMAIL PROTECTED] wrote: On Sep 24, 2008, at 2:40 PM, Jacques Basaldúa [EMAIL PROTECTED] wr ote: Therefore, the variance of the normal that best approximates the distribution of both RAVE and wins/(wins + losses) is the same n·p·(1-p) See above, it's slightly different. If this is true, the variance you are measuring from the samples does not contain any information about the precision of the estimators. If someone understands this better, please explain it to the list. This will get covered in my next revision. A proper discussion is too much to type with my thumb... My paper writing time is less than I hadhiped, so here's a quick and dirty answer. For a fixed win rate, the probabilities of a specific number of wins and losses follows the binomial distribution. That distribution keeps p (probability of winning) constant and the number of observed wins and losses variable. When trying to reverse this process, the wins and losses are kept constant and p varies. Essentially prob(p=x) is proportional to (x^wins)(1-x)^losses. This is a Beta distribution with known mean, mode, variance, etc... It's these values which should be used for approximating the win rate estimator as a normal distion. Does that makes sense? I've glossed over a very important detail when reversing. Bayes Theorem requires some extra a priori information. My preferred handling alters the reversed equation's exponents a bit but the basic conclusion (of a beta distribution) is the same.___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch (Jason House's paper)
On Fri, Sep 26, 2008 at 9:29 AM, Jason House [EMAIL PROTECTED] wrote: Sent from my iPhone On Sep 24, 2008, at 5:16 PM, Jason House [EMAIL PROTECTED] wrote: On Sep 24, 2008, at 2:40 PM, Jacques Basaldúa [EMAIL PROTECTED] wrote: Therefore, the variance of the normal that best approximates the distribution of both RAVE and wins/(wins + losses) is the same n·p·(1-p) See above, it's slightly different. If this is true, the variance you are measuring from the samples does not contain any information about the precision of the estimators. If someone understands this better, please explain it to the list. This will get covered in my next revision. A proper discussion is too much to type with my thumb... My paper writing time is less than I hadhiped, so here's a quick and dirty answer. For a fixed win rate, the probabilities of a specific number of wins and losses follows the binomial distribution. That distribution keeps p (probability of winning) constant and the number of observed wins and losses variable. When trying to reverse this process, the wins and losses are kept constant and p varies. Essentially prob(p=x) is proportional to (x^wins)(1-x)^losses. This is a Beta distribution with known mean, mode, variance, etc... It's these values which should be used for approximating the win rate estimator as a normal distion. Does that makes sense? I've glossed over a very important detail when reversing. Bayes Theorem requires some extra a priori information. My preferred handling alters the reversed equation's exponents a bit but the basic conclusion (of a beta distribution) is the same. Maybe I can say it a little more precisely. Before we have collected any data, let's use a uniform prior for p. After we sample the move a number of times, we obtain w wins and l losses. Bayes's theorem tells us that the posterior probability distribution is a beta distribution B(w+1,l+1) (see http://en.wikipedia.org/wiki/Beta_distribution for details). An implication of this is that the expected value of p after w wins and l losses is (w+1)/(w+l+2). This is the same as initializing w=l=1 before you have any information and then using w/(w+l) as your winning rate, which some people have done intuitively, but it's clear that it's not just a kludge. I'll use the letter r for the value of the winning rate. The estimate of the variance is (w+1)*(l+1)/((w+l+2)^2*(w+l+3)), which is r*(1-r)/(w+l+3). The simple UCB formula uses an estimate of the variance that is simply 1/visits, so perhaps one should modify the formula by multiplying that estimate by r*(1-r), which means that the variance is smaller in positions that look like clear victories for one side. I don't know if this makes any difference in practice, but I doubt it. Álvaro. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch (Jason House's paper)
Sent from my iPhone On Sep 27, 2008, at 10:14 AM, Álvaro Begué [EMAIL PROTECTED] wrote: On Fri, Sep 26, 2008 at 9:29 AM, Jason House [EMAIL PROTECTED] wrote: Sent from my iPhone On Sep 24, 2008, at 5:16 PM, Jason House [EMAIL PROTECTED] wrote: On Sep 24, 2008, at 2:40 PM, Jacques Basaldúa [EMAIL PROTECTED] wrote: Therefore, the variance of the normal that best approximates the distribution of both RAVE and wins/(wins + losses) is the same n·p·(1-p) See above, it's slightly different. If this is true, the variance you are measuring from the samples does not contain any information about the precision of the estimators. If someone understands this better, please explain it to the list. This will get covered in my next revision. A proper discussion is too much to type with my thumb... My paper writing time is less than I hadhiped, so here's a quick and dirty answer. For a fixed win rate, the probabilities of a specific number of wins and losses follows the binomial distribution. That distribution keeps p (probability of winning) constant and the number of observed wins and losses variable. When trying to reverse this process, the wins and losses are kept constant and p varies. Essentially prob(p=x) is proportional to (x^wins)(1- x)^losses. This is a Beta distribution with known mean, mode, variance, etc... It's these values which should be used for approximating the win rate estimator as a normal distion. Does that makes sense? I've glossed over a very important detail when reversing. Bayes Theorem requires some extra a priori information. My preferred handling alters the reversed equation's exponents a bit but the basic conclusion (of a beta distribution) is the same. Maybe I can say it a little more precisely. Before we have collected any data, let's use a uniform prior for p. After we sample the move a number of times, we obtain w wins and l losses. Bayes's theorem tells us that the posterior probability distribution is a beta distribution B(w+1,l+1) (see http://en.wikipedia.org/wiki/Beta_distribution for details). When I originally posted about this stuff and modifying the UCB formula, the uniform prior was a major sticking point for people. It is my preferred handling when no domain knowledge/heuristics are used. An implication of this is that the expected value of p after w wins and l losses is (w+1)/(w+l+2). This is the same as initializing w=l=1 before you have any information and then using w/(w+l) as your winning rate, which some people have done intuitively, but it's clear that it's not just a kludge. I'll use the letter r for the value of the winning rate. The estimate of the variance is (w+1)*(l+1)/((w+l+2)^2*(w+l+3)), which is r*(1-r)/(w+l+3). The simple UCB formula uses an estimate of the variance that is simply 1/visits, so perhaps one should modify the formula by multiplying that estimate by r*(1-r), which means that the variance is smaller in positions that look like clear victories for one side. I don't know if this makes any difference in practice, but I doubt it. The UCB1-Tuned formula aimed at a similar effect. I know the MoGo team abandoned it because they didn't see a strength gain, just added complexity. In fact, they don't even use UCB's at all anymore. Also, note that when initializing wins and losses to 1, n=wins+losses and p=wins/n, the variance becomes p(1-p)/(n+1) which is very close to a nieve binomial distribution of p(1-p)/n with win and loss initialization. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch
On Sep 25, 2008, at 12:12 AM, David Fotland [EMAIL PROTECTED] games.com wrote: This is an interesting idea, but do you have any actual results? If you implement this kind of rave formula do you get a stronger program? I don't have enough data to say anything except that it did not have a large (100 ELO+) change. I switched to this method just before the July KGS tournament. Both the old and new versions were untuned, but based on parameters I've heard on this list. My testing on CGOS is tough to draw conclusions from. David -Original Message- From: [EMAIL PROTECTED] [mailto:computer-go- [EMAIL PROTECTED] On Behalf Of Jason House Sent: Wednesday, September 24, 2008 4:34 AM To: computer-go Subject: Re: [computer-go] MoGo v.s. Kim rematch On Tue, 2008-09-23 at 18:08 -0300, Douglas Drumond wrote: Attached is a quick write up of what I was talking about with some math. PS: Any tips on cleanup and making it a mini publication would be appreciated. I've never published a paper before. Would this be too small? Better add an abstract, but what I missed most was bibliography. Ask and you shall receive :) Actually, I spent most of my free time learning Tex/Lyx, so there are very few changes in this version. I'm out of time for a while, so I figured I'd just share what I have so far. []'s Douglas Drumond - Computer Engineering FEEC/IC - Unicamp ___ 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] MoGo v.s. Kim rematch (Jason House's paper)
The approach of this paper is to treat all win rate estimations as independent estimators with additive white Gaussian noise. Have you tried if that works? (As Łukasz Lew wrote experimental setup would be useful) I guess there may be a flaw in your idea, but I am not a specialist. I will try to explain it. If it wasn't for the fact that the tree is learning, the probability of a playout through a node to win would be constant each time the node is visited. This is, of course, a simplification because the tree does learn, but, at least between playouts that are not very distant in time, it is true. So my argument holds to some (I guess, much) extent. The same applies to the RAVE estimator which is also the result of counting wins (assume P(win|that move) = constant) and dividing by some appropriate sample size. Therefore, these estimators follow a binomial distribution. It does converge to the normal, but with some fundamental caveat: Unlike the normal in which mean an variance are independent, in this case the variance is a function of p. The variance of the binomial = n·p·(1-p) is a _function of p_. Therefore, the variance of the normal that best approximates the distribution of both RAVE and wins/(wins + losses) is the same n·p·(1-p) If this is true, the variance you are measuring from the samples does not contain any information about the precision of the estimators. If someone understands this better, please explain it to the list. Jacques.* * ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] MoGo v.s. Kim rematch
This is an interesting idea, but do you have any actual results? If you implement this kind of rave formula do you get a stronger program? David -Original Message- From: [EMAIL PROTECTED] [mailto:computer-go- [EMAIL PROTECTED] On Behalf Of Jason House Sent: Wednesday, September 24, 2008 4:34 AM To: computer-go Subject: Re: [computer-go] MoGo v.s. Kim rematch On Tue, 2008-09-23 at 18:08 -0300, Douglas Drumond wrote: Attached is a quick write up of what I was talking about with some math. PS: Any tips on cleanup and making it a mini publication would be appreciated. I've never published a paper before. Would this be too small? Better add an abstract, but what I missed most was bibliography. Ask and you shall receive :) Actually, I spent most of my free time learning Tex/Lyx, so there are very few changes in this version. I'm out of time for a while, so I figured I'd just share what I have so far. []'s Douglas Drumond - Computer Engineering FEEC/IC - Unicamp ___ 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] MoGo v.s. Kim rematch
On 22, Sep 2008, at 10:50 PM, Hideki Kato wrote: David Doshay: [EMAIL PROTECTED]: It was 800, just like last time, but the networking had been upgraded from ethernet to infiniband. Olivier said that this should have been a good improvement because he felt that communication overhead was significant. Really previous Huygens used Ethernet? It's hard to believe... Hideki I thought so as well, but Olivier wrote to me: Begin forwarded message: From: Olivier Teytaud [EMAIL PROTECTED] Date: 6, September 2008 2:07:42 AM PDT To: David Doshay [EMAIL PROTECTED] Subject: Re: [computer-go] yet a mogo vs human game Hi David, ... We will have at least the same number of cores, probably more, and we will very likely have a better hardware - the infiniband network should be available, and this makes a big difference. ... Best regards, Olivier So, perhaps Huygens has both and they were not using it last time, or maybe they brought Huygens up with E-net and then upgraded. But Mogo did not use it for the Portland exhibition, but did use infiniband for the rematch. Cheers, David ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch
David, I've found a description that Infiniband was improved from 2 x 4X IB (20 Gbps) to 8 x 8X IB (160 Gbps) on Jun 2008 at the bottom of 6th page of a pdf about Huygens system: https://www.os3.nl/_media/2007-2008/courses/inr/week7/sne_20080320_walter.pdf I guess that is the better hardware Olivier wrote. Hideki David Doshay: [EMAIL PROTECTED]: On 22, Sep 2008, at 10:50 PM, Hideki Kato wrote: David Doshay: [EMAIL PROTECTED]: It was 800, just like last time, but the networking had been upgraded from ethernet to infiniband. Olivier said that this should have been a good improvement because he felt that communication overhead was significant. Really previous Huygens used Ethernet? It's hard to believe... Hideki I thought so as well, but Olivier wrote to me: Begin forwarded message: From: Olivier Teytaud [EMAIL PROTECTED] Date: 6, September 2008 2:07:42 AM PDT To: David Doshay [EMAIL PROTECTED] Subject: Re: [computer-go] yet a mogo vs human game Hi David, ... We will have at least the same number of cores, probably more, and we will very likely have a better hardware - the infiniband network should be available, and this makes a big difference. ... Best regards, Olivier So, perhaps Huygens has both and they were not using it last time, or maybe they brought Huygens up with E-net and then upgraded. But Mogo did not use it for the Portland exhibition, but did use infiniband for the rematch. Cheers, David ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ -- [EMAIL PROTECTED] (Kato) ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch
On Mon, Sep 22, 2008 at 1:21 PM, Łukasz Lew [EMAIL PROTECTED] wrote: Hi, On Mon, Sep 22, 2008 at 17:58, Jason House [EMAIL PROTECTED] wrote: On Sep 22, 2008, at 7:59 AM, Magnus Persson [EMAIL PROTECTED] wrote: The results of the math are most easilly expressed in terms of inverse variance (iv=1/variance) Combined mean = sum( mean * iv ) Combined iv = sum( iv ) I'll try to do a real write-up if anyone is interested. I am very interested. :) Lukasz Attached is a quick write up of what I was talking about with some math. PS: Any tips on cleanup and making it a mini publication would be appreciated. I've never published a paper before. Would this be too small? RAVE.doc Description: MS-Word document ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch
Attached is a quick write up of what I was talking about with some math. PS: Any tips on cleanup and making it a mini publication would be appreciated. I've never published a paper before. Would this be too small? Better add an abstract, but what I missed most was bibliography. []'s Douglas Drumond - Computer Engineering FEEC/IC - Unicamp ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch
It was 800, just like last time, but the networking had been upgraded from ethernet to infiniband. Olivier said that this should have been a good improvement because he felt that communication overhead was significant. I thought Olivier had previously said there was very little overhead. E.g.: http://computer-go.org/pipermail/computer-go/2008-May/015068.html http://www.mail-archive.com/computer-go@computer-go.org/msg07953.html I took this 95% to mean that giving Mogo 760 (800x0.95) times the thinking time on a single core would be the same strength as Mogo on 800 cores. Assuming a rule of thumb that doubling the playouts is worth one rank (*), increasing network speed will surely have very little effect on the strength? Darren *: That is from memory of discussion here; I hope somebody will correct me with a more accurate thumb. We will have at least the same number of cores, probably more, and we will very likely have a better hardware - the infiniband network should be available, and this makes a big difference. -- Darren Cook, Software Researcher/Developer http://dcook.org/mlsn/ (English-Japanese-German-Chinese-Arabic open source dictionary/semantic network) http://dcook.org/work/ (About me and my work) http://dcook.org/blogs.html (My blogs and articles) ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch
Quoting Mark Boon [EMAIL PROTECTED]: Playing out that fake ladder in the first game meant an instant loss. Surprising. And embarassing. Any information on the number of processors used? The interesting question is if there is a silly bug or something more sophisticated. I have struggled with ladders in Valkyria and it is often really hard to tell what causes these problems. In Leksand most games on 19x19 where lost in a ways similar to the recent mogo game. I could not find an obvious problem with the playouts at least not in terms of an easily fixable bug. Note that Valkyria reads out 99% of all ladders correctly both in the tree and in the playouts. What I realized was that AMAF in combination with heavy playouts causes some serious biases, for some kinds of very bad moves such that AMAF completely misevaluate them. In the case of the ladders the heavy playouts of Valkyria correctly prunes playing out ladders for the loser. But sometimes in the playouts the ladder is broken and after that there is a chance that the stones escape anyway. This means that almost always when the escaping move is played it is a good move! Thus AMAF will assign a very good score to this move My solutions to this was simply to turn off AMAF-eval for all shapes commonly misevaluated for ladders. But I think this problem is true for many shapes in general. What makes ladders special is that the problem repeats it self and the effect get stronger and thus even more likely the larger the ladder gets. I think a better solution would be to modify AMAF in some way to avoid these problems, or perhaps change the playouts in a way to balance the problem. Does anyone know something to do about it or have any ideas? -Magnus ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch
I'm curious about a couple of things in particular. Is this a bug and how much time would be required for Mogo to have played the correct move if it wasn't. Of course I'm also interested in ways to solve this with less deep searches or better play-outs. - Don On Mon, 2008-09-22 at 13:59 +0200, Magnus Persson wrote: Quoting Mark Boon [EMAIL PROTECTED]: Playing out that fake ladder in the first game meant an instant loss. Surprising. And embarassing. Any information on the number of processors used? The interesting question is if there is a silly bug or something more sophisticated. I have struggled with ladders in Valkyria and it is often really hard to tell what causes these problems. In Leksand most games on 19x19 where lost in a ways similar to the recent mogo game. I could not find an obvious problem with the playouts at least not in terms of an easily fixable bug. Note that Valkyria reads out 99% of all ladders correctly both in the tree and in the playouts. What I realized was that AMAF in combination with heavy playouts causes some serious biases, for some kinds of very bad moves such that AMAF completely misevaluate them. In the case of the ladders the heavy playouts of Valkyria correctly prunes playing out ladders for the loser. But sometimes in the playouts the ladder is broken and after that there is a chance that the stones escape anyway. This means that almost always when the escaping move is played it is a good move! Thus AMAF will assign a very good score to this move My solutions to this was simply to turn off AMAF-eval for all shapes commonly misevaluated for ladders. But I think this problem is true for many shapes in general. What makes ladders special is that the problem repeats it self and the effect get stronger and thus even more likely the larger the ladder gets. I think a better solution would be to modify AMAF in some way to avoid these problems, or perhaps change the playouts in a way to balance the problem. Does anyone know something to do about it or have any ideas? -Magnus ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ signature.asc Description: This is a digitally signed message part ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch
Consider this as tentative, since I heard it about 3rd-hand, but I believe the number of processors used to have been 3000. Congratulations to the Mogo team; good luck improving your program to deal with the ladder and life-and-death issues. Looking forward to further information! I have always wondered if AMAF is a feature or a bug. There are many situations where the order of moves is crucial; A before B wins, B before A loses; ladders are a classic example where the ordering of moves is utterly important. AMAF seems to assume that order doesn't matter. Of course, there are many other positions where this assumption is true; that is why it sometimes yields an improvement in processing speed, but it seems risky. Ladders are also a classic case where two patterns can look very similar, but be very different. When you capture a ladder, you are in a very good position. But if the stones under attack have just one extra liberty, the position may look like a ladder, but your target will escape, and your stones will be full of cutting points; the proper evaluation for that position would be much harsher. More generally, whenever I see a Monte Carlo program lose, it is usually a semeai where being one liberty behind or one ahead makes all the difference. We call these capturing races in English for a reason; being ahead or behind by one liberty matters a great deal. To make life interesting, there are loose ladder constructs where an extra liberty does not help the fleeing stones; they still get corraled and captured. These corner cases are tough, but many games hinge on correctly reading out the exact consequences of life-and-death struggles. Terry McIntyre [EMAIL PROTECTED] Go is very hard. The more I learn about it, the less I know. -Jie Li, 9 dan On Mon, 2008-09-22 at 13:59 +0200, Magnus Persson wrote: Quoting Mark Boon : Playing out that fake ladder in the first game meant an instant loss. Surprising. And embarassing. Any information on the number of processors used? ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch
I think AMAF is a feature not a bug. It's only a matter of how judiciously it's applied. Also, almost any evaluation feature in a game playing program is a bug - meaning it is an imperfect approximation of what you really want. Of course it could turn out that AMAF got them in trouble in this game. The Mogo team will probably analyze the reason for the problem.But as long as they are playing strong professional players they are going to have something to debug and analyze! - Don On Mon, 2008-09-22 at 06:06 -0700, terry mcintyre wrote: Consider this as tentative, since I heard it about 3rd-hand, but I believe the number of processors used to have been 3000. Congratulations to the Mogo team; good luck improving your program to deal with the ladder and life-and-death issues. Looking forward to further information! I have always wondered if AMAF is a feature or a bug. There are many situations where the order of moves is crucial; A before B wins, B before A loses; ladders are a classic example where the ordering of moves is utterly important. AMAF seems to assume that order doesn't matter. Of course, there are many other positions where this assumption is true; that is why it sometimes yields an improvement in processing speed, but it seems risky. Ladders are also a classic case where two patterns can look very similar, but be very different. When you capture a ladder, you are in a very good position. But if the stones under attack have just one extra liberty, the position may look like a ladder, but your target will escape, and your stones will be full of cutting points; the proper evaluation for that position would be much harsher. More generally, whenever I see a Monte Carlo program lose, it is usually a semeai where being one liberty behind or one ahead makes all the difference. We call these capturing races in English for a reason; being ahead or behind by one liberty matters a great deal. To make life interesting, there are loose ladder constructs where an extra liberty does not help the fleeing stones; they still get corraled and captured. These corner cases are tough, but many games hinge on correctly reading out the exact consequences of life-and-death struggles. Terry McIntyre [EMAIL PROTECTED] Go is very hard. The more I learn about it, the less I know. -Jie Li, 9 dan On Mon, 2008-09-22 at 13:59 +0200, Magnus Persson wrote: Quoting Mark Boon : Playing out that fake ladder in the first game meant an instant loss. Surprising. And embarassing. Any information on the number of processors used? ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ signature.asc Description: This is a digitally signed message part ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] MoGo v.s. Kim rematch
AMAF certainly helps to do move ordering when there is little other information. With good prior heuristics or enough actual playouts, it should not be weighted very highly. AMAF finds good moves, but it often bias heavily for or against moves. In ManyFaces, AMAF (actually RAVE) is worth between 5% and 10% wins against gnugo. I've seen similar ladder problems, and it is not AMAF, it's caused by the playouts, when they can't read ladders. It's easy to add various hacks to prevent playing out simple ladders, but the one in this game had an extra liberty (if I remember correctly). A general solution is a little tricky. David -Original Message- From: [EMAIL PROTECTED] [mailto:computer-go- [EMAIL PROTECTED] On Behalf Of Don Dailey Sent: Monday, September 22, 2008 6:23 AM To: computer-go Subject: Re: [computer-go] MoGo v.s. Kim rematch I think AMAF is a feature not a bug. It's only a matter of how judiciously it's applied. Also, almost any evaluation feature in a game playing program is a bug - meaning it is an imperfect approximation of what you really want. Of course it could turn out that AMAF got them in trouble in this game. The Mogo team will probably analyze the reason for the problem.But as long as they are playing strong professional players they are going to have something to debug and analyze! - Don On Mon, 2008-09-22 at 06:06 -0700, terry mcintyre wrote: Consider this as tentative, since I heard it about 3rd-hand, but I believe the number of processors used to have been 3000. Congratulations to the Mogo team; good luck improving your program to deal with the ladder and life-and-death issues. Looking forward to further information! I have always wondered if AMAF is a feature or a bug. There are many situations where the order of moves is crucial; A before B wins, B before A loses; ladders are a classic example where the ordering of moves is utterly important. AMAF seems to assume that order doesn't matter. Of course, there are many other positions where this assumption is true; that is why it sometimes yields an improvement in processing speed, but it seems risky. Ladders are also a classic case where two patterns can look very similar, but be very different. When you capture a ladder, you are in a very good position. But if the stones under attack have just one extra liberty, the position may look like a ladder, but your target will escape, and your stones will be full of cutting points; the proper evaluation for that position would be much harsher. More generally, whenever I see a Monte Carlo program lose, it is usually a semeai where being one liberty behind or one ahead makes all the difference. We call these capturing races in English for a reason; being ahead or behind by one liberty matters a great deal. To make life interesting, there are loose ladder constructs where an extra liberty does not help the fleeing stones; they still get corraled and captured. These corner cases are tough, but many games hinge on correctly reading out the exact consequences of life-and-death struggles. Terry McIntyre [EMAIL PROTECTED] Go is very hard. The more I learn about it, the less I know. -Jie Li, 9 dan On Mon, 2008-09-22 at 13:59 +0200, Magnus Persson wrote: Quoting Mark Boon : Playing out that fake ladder in the first game meant an instant loss. Surprising. And embarassing. Any information on the number of processors used? ___ 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] MoGo v.s. Kim rematch
On Sep 22, 2008, at 7:59 AM, Magnus Persson [EMAIL PROTECTED] wrote: In the case of the ladders the heavy playouts of Valkyria correctly prunes playing out ladders for the loser. But sometimes in the playouts the ladder is broken and after that there is a chance that the stones escape anyway. This means that almost always when the escaping move is played it is a good move! Thus AMAF will assign a very good score to this move My solutions to this was simply to turn off AMAF-eval for all shapes commonly misevaluated for ladders. But I think this problem is true for many shapes in general. What makes ladders special is that the problem repeats it self and the effect get stronger and thus even more likely the larger the ladder gets. I think a better solution would be to modify AMAF in some way to avoid these problems, or perhaps change the playouts in a way to balance the problem. Does anyone know something to do about it or have any ideas? My RAVE formulation includes a per-move parameter for RAVE confidence. This allows heuristics to fix situations like above. Sadly, my bot isn't mature enough to take advantage yet :( The concept I used for the derivation is simple. I treat everything as gaussian estimators. It's easy to find the max of the distribution. I then use the same trick as bayeselo to estimate variance. I then add a Gaussian noise term to represent RAVE bias. The results of the math are most easilly expressed in terms of inverse variance (iv=1/variance) Combined mean = sum( mean * iv ) Combined iv = sum( iv ) I'll try to do a real write-up if anyone is interested. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch
It was 800, just like last time, but the networking had been upgraded from ethernet to infiniband. Olivier said that this should have been a good improvement because he felt that communication overhead was significant. Cheers, David On 22, Sep 2008, at 6:06 AM, terry mcintyre wrote: Consider this as tentative, since I heard it about 3rd-hand, but I believe the number of processors used to have been 3000. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch
David Doshay: [EMAIL PROTECTED]: It was 800, just like last time, but the networking had been upgraded from ethernet to infiniband. Olivier said that this should have been a good improvement because he felt that communication overhead was significant. Really previous Huygens used Ethernet? It's hard to believe... Hideki Cheers, David On 22, Sep 2008, at 6:06 AM, terry mcintyre wrote: Consider this as tentative, since I heard it about 3rd-hand, but I believe the number of processors used to have been 3000. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ -- [EMAIL PROTECTED] (Kato) ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch
Anyone knows the result, or better the game sgf? On Sat, Sep 6, 2008 at 6:57 AM, Don Dailey [EMAIL PROTECTED] wrote: Great news! Look forward to seeing it happen. I hope Mogo has some great hardware. - Don On Fri, 2008-09-05 at 15:54 -0700, David Doshay wrote: MoGo and Myungwan Kim will hold an exhibition rematch at the Cotsen Open on Saturday September 20. The exhibition will start at about 5pm Pacific Daylight time. As probably known by all on this list, MoGo won the last game, held at the US Go Congress in Portland Oregon, when it was given a 9 stone handicap and played with a one hour time limit. At this time the expected handicap will be 7, and it is not clear if there will be one game or two. It is not known at this time how many cores MoGo will be running on. Mr Kim has asked for MoGo to be given 90 minutes because he saw how much the increase in time from 15 minutes to one hour increased its playing strength. Mr. Kim has also asked that there be only one or at most 2 blitz games at the start. The MoGo team also wants to have some 9x9 games, but Mr Kim does not feel familiar enough with 9x9 to play those games, but he is searching for an alternate strong player. MoGo has some new features for 9x9, and the team is anxious to see the newest code in action. Cheers, David ___ 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] MoGo v.s. Kim rematch
There is a discussion at godiscussions http://www.godiscussions.com/forum/showthread.php?t=7154 ?The sgf's for the two games played are on page 2 Dave Hillis -Original Message- From: mingwu [EMAIL PROTECTED] To: [EMAIL PROTECTED]; computer-go computer-go@computer-go.org Sent: Sun, 21 Sep 2008 9:22 pm Subject: Re: [computer-go] MoGo v.s. Kim rematch Anyone knows the result, or better the game sgf? On Sat, Sep 6, 2008 at 6:57 AM, Don Dailey [EMAIL PROTECTED] wrote: Great news! ? Look forward to seeing it happen. ?I hope Mogo has some great hardware. - Don On Fri, 2008-09-05 at 15:54 -0700, David Doshay wrote: MoGo and Myungwan Kim will hold an exhibition rematch at the Cotsen Open on Saturday September 20. The exhibition will start at about 5pm Pacific Daylight time. As probably known by all on this list, MoGo won the last game, held at the US Go Congress in Portland Oregon, when it was given a 9 stone handicap and played with a one hour time limit. At this time the expected handicap will be 7, and it is not clear if there will be one game or two. It is not known at this time how many cores MoGo will be running on. Mr Kim has asked for MoGo to be given 90 minutes because he saw how much the increase in time from 15 minutes to one hour increased its playing strength. Mr. Kim has also asked that there be only one or at most 2 blitz games at the start. The MoGo team also wants to have some 9x9 games, but Mr Kim does not feel familiar enough with 9x9 to play those games, but he is searching for an alternate strong player. MoGo has some new features for 9x9, and the team is anxious to see the newest code in action. Cheers, David ___ 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/
[computer-go] MoGo v.s. Kim rematch
MoGo and Myungwan Kim will hold an exhibition rematch at the Cotsen Open on Saturday September 20. The exhibition will start at about 5pm Pacific Daylight time. As probably known by all on this list, MoGo won the last game, held at the US Go Congress in Portland Oregon, when it was given a 9 stone handicap and played with a one hour time limit. At this time the expected handicap will be 7, and it is not clear if there will be one game or two. It is not known at this time how many cores MoGo will be running on. Mr Kim has asked for MoGo to be given 90 minutes because he saw how much the increase in time from 15 minutes to one hour increased its playing strength. Mr. Kim has also asked that there be only one or at most 2 blitz games at the start. The MoGo team also wants to have some 9x9 games, but Mr Kim does not feel familiar enough with 9x9 to play those games, but he is searching for an alternate strong player. MoGo has some new features for 9x9, and the team is anxious to see the newest code in action. Cheers, David ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo v.s. Kim rematch
Great news! Look forward to seeing it happen. I hope Mogo has some great hardware. - Don On Fri, 2008-09-05 at 15:54 -0700, David Doshay wrote: MoGo and Myungwan Kim will hold an exhibition rematch at the Cotsen Open on Saturday September 20. The exhibition will start at about 5pm Pacific Daylight time. As probably known by all on this list, MoGo won the last game, held at the US Go Congress in Portland Oregon, when it was given a 9 stone handicap and played with a one hour time limit. At this time the expected handicap will be 7, and it is not clear if there will be one game or two. It is not known at this time how many cores MoGo will be running on. Mr Kim has asked for MoGo to be given 90 minutes because he saw how much the increase in time from 15 minutes to one hour increased its playing strength. Mr. Kim has also asked that there be only one or at most 2 blitz games at the start. The MoGo team also wants to have some 9x9 games, but Mr Kim does not feel familiar enough with 9x9 to play those games, but he is searching for an alternate strong player. MoGo has some new features for 9x9, and the team is anxious to see the newest code in action. Cheers, David ___ 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] MoGo beats pro: The website
Dear all, There were details that were unclear about the victory of MoGo. Hence I created a website to gather useful information about this game: http://www.cs.unimaas.nl/g.chaslot/muyungwan-mogo/ Cheers, Guillaume Message d'origine De: [EMAIL PROTECTED] de la part de Sylvain Gelly Date: mer. 8/13/2008 19:54 À: computer-go Objet : Re: [computer-go] What was the specific design of the Mogo versionwhich beat the pro... C++ on linux (with a port on windows using cygwin libraries for the binary release) Sylvain 2008/8/13 steve uurtamo [EMAIL PROTECTED] And what language/platform is Mogo written in; C/C++, Java, Assembly, PHP, etc.? This made coffee spray out of my nose (PHP). I think that C is most likely, based upon how they parallelized it. Did you read the list posting that mentioned (briefly) how they scaled it up? s. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ winmail.dat___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo beats pro: The website
If this is aimed at clearing up ambiguity, you should state which way the handicap was given. Sent from my iPhone On Aug 13, 2008, at 2:08 PM, Chaslot G (MICC) [EMAIL PROTECTED] wrote: Dear all, There were details that were unclear about the victory of MoGo. Hence I created a website to gather useful information about this game: http://www.cs.unimaas.nl/g.chaslot/muyungwan-mogo/ Cheers, Guillaume Message d'origine De: [EMAIL PROTECTED] de la part de Sylvain Gelly Date: mer. 8/13/2008 19:54 À: computer-go Objet : Re: [computer-go] What was the specific design of the Mogo versionwhich beat the pro... C++ on linux (with a port on windows using cygwin libraries for the binary release) Sylvain 2008/8/13 steve uurtamo [EMAIL PROTECTED] And what language/platform is Mogo written in; C/C++, Java, Assembly, PHP, etc.? This made coffee spray out of my nose (PHP). I think that C is most likely, based upon how they parallelized it. Did you read the list posting that mentioned (briefly) how they scaled it up? s. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ winmail.dat ___ 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] MoGo beats pro: The website
From: Jason House [EMAIL PROTECTED] If this is aimed at clearing up ambiguity, you should state which way the handicap was given. Oops! Now I need to clean off my keyboard! rotflmao! Mmmm, we already have a hotly-contested estimate that computer programs will play pros on an even basis in ten years time. Shall we now open the pool for computer offers pro a 9 stone handicap and wins? On Aug 13, 2008, at 2:08 PM, Chaslot G (MICC) [EMAIL PROTECTED] wrote: Dear all, There were details that were unclear about the victory of MoGo. Hence I created a website to gather useful information about this game: http://www.cs.unimaas.nl/g.chaslot/muyungwan-mogo/ Cheers, Guillaume ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] mogo beats pro!
From: Bob Hearn [EMAIL PROTECTED] Now, my question. Sorry if this has already been beaten to death here. After the match, one of the MoGo programmers mentioned that doubling the computation led to a 63% win rate against the baseline version, and that so far this scaling seemed to continue as computation power increased. So -- quick back-of-the-envelope calculation, tell me where I am wrong. 63% win rate = about half a stone advantage in go. So we need 4x processing power to increase by a stone. At the current rate of Moore's law, that's about 4 years. Kim estimated that the game with MoGo would be hard at 8 stones. That suggests that in 32 years a supercomputer comparable to the one that played in this match would be as strong as Kim. This calculation is optimistic in assuming that you can meaningfully scale the 63% win rate indefinitely, especially when measuring strength against other opponents, and not a weaker version of itself. It's also pessimistic in assuming there will be no improvement in the Monte Carlo technique. But still, 32 years seems like a surprisingly long time, much longer than the 10 years that seems intuitively reasonable. Naively, it would seem that improvements in the Monte Carlo algorithms could gain some small number of stones in strength for fixed computation, but that would just shrink the 32 years by maybe a decade. How do others feel about this? I guess I should also go on record as believing that if it really does take 32 years, we *will* have general-purpose AI before then. I suspect that Mogo -- good as it is -- is far from being the optimal algorithm. In ten years time new methods will emerge which should yield considerable improvements. In addition, the 800-core supercomputer used was not today's state of the art; the Mogo team almost obtained a 3000-core supercomputer for this exhibition, which would be nearly 4x as large; as Computer Go becomes more exciting, we may be able to borrow still more impressive hardware -- current state-of-the-art is 65k or even 128k processors. Third, the 32 year figure is highly sensitive to one's expectation of Moore's Law. A doubling every 18 months would be a quadrupling every 36 months, which is three years; this factor alone shrinks the 32 years to 24. We may see a faster rate of growth - GPUs have been improving faster than general-purpose CPUs, and the coming multicore processors may have more in common with GPUs than with previous generations of X86 cores -- we may revert to simpler RISC cores, which use less silicon. In short, reaching the top of the pyramid would be a thousand-fold improvement in processing power -- about 4 to the 4th power, or half way to the goal. During the same period, the petaflops race and Moore's Law would continue to increase the power of the Top 500. Stir in some algorithmic improvements, and we should be within range in something closer to ten, not 32 years. If general purpose AI means an AI which can solve every problem at the expert level, that is probably not a prerequisite for solving one problem at an expert level. We're not asking for a program which can skillfully play a teaching game against a weaker player, as a human pro would, nor are we asking that it be able to dance the salsa; it just needs to beat a pro in an even game. We have just barely started optimizing the search. What do humans know that computers don't? How do pros manage to play well without the ability to examine trillions of playouts? ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] mogo beats pro!
David Doshay wrote: As an aside, the pro in question won the US Open, so comments about him being a weak pro seem inappropriate. I spoke with him a number of times, and I firmly believe that he took the match as seriously as any other public exhibition of his skill that involves handicap stones for the opponent. He has an open mind about computer Go, unlike some other pro players I talked to here at the congress. Kim did state before the match that in his opinion computers would never be as strong as the best humans. I don't believe he was asked afterward whether he had any reason to change that opinion. After the banquet last night, I was talking to Peter Drake when Kim walked up and started asking questions about how MoGo played go. Peter explained very well, but I'm not sure he completely understood. BTW, David, I also pointed out to Chris Garlock that you'd been misquoted, shortly after the story went up on the AGA website, but he didn't reply. Also BTW, let me introduce myself to the list and ask a question. I'm a 2D go player, also an AI researcher affiliated with Dartmouth. I did my Ph.D. at MIT on games and puzzles. However, I never seriously worked on computer go, because I was always convinced go was AI- complete -- that we would have strong go programs when we had general- purpose AI, and not before. Mostly my current work is on general- purpose AI heavily inspired by neuroscience. However, with the advent of the Monte Carlo programs I'm about ready to change my mind. I'm tempted to try to work in the area and see whether I can contribute anything. Now, my question. Sorry if this has already been beaten to death here. After the match, one of the MoGo programmers mentioned that doubling the computation led to a 63% win rate against the baseline version, and that so far this scaling seemed to continue as computation power increased. So -- quick back-of-the-envelope calculation, tell me where I am wrong. 63% win rate = about half a stone advantage in go. So we need 4x processing power to increase by a stone. At the current rate of Moore's law, that's about 4 years. Kim estimated that the game with MoGo would be hard at 8 stones. That suggests that in 32 years a supercomputer comparable to the one that played in this match would be as strong as Kim. This calculation is optimistic in assuming that you can meaningfully scale the 63% win rate indefinitely, especially when measuring strength against other opponents, and not a weaker version of itself. It's also pessimistic in assuming there will be no improvement in the Monte Carlo technique. But still, 32 years seems like a surprisingly long time, much longer than the 10 years that seems intuitively reasonable. Naively, it would seem that improvements in the Monte Carlo algorithms could gain some small number of stones in strength for fixed computation, but that would just shrink the 32 years by maybe a decade. How do others feel about this? I guess I should also go on record as believing that if it really does take 32 years, we *will* have general-purpose AI before then. Bob Hearn ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] mogo beats pro!
On Sun, 2008-08-10 at 11:37 -0700, Bob Hearn wrote: Now, my question. Sorry if this has already been beaten to death here. After the match, one of the MoGo programmers mentioned that doubling the computation led to a 63% win rate against the baseline version, and that so far this scaling seemed to continue as computation power increased. So -- quick back-of-the-envelope calculation, tell me where I am wrong. 63% win rate = about half a stone advantage in go. So we need 4x processing power to increase by a stone. At the current rate of Moore's law, that's about 4 years. Kim estimated that the game with MoGo would be hard at 8 stones. That suggests that in 32 years a supercomputer comparable to the one that played in this match would be as strong as Kim. This calculation is optimistic in assuming that you can meaningfully scale the 63% win rate indefinitely, especially when measuring strength against other opponents, and not a weaker version of itself. It's also pessimistic in assuming there will be no improvement in the Monte Carlo technique. But still, 32 years seems like a surprisingly long time, much longer than the 10 years that seems intuitively reasonable. Naively, it would seem that improvements in the Monte Carlo algorithms could gain some small number of stones in strength for fixed computation, but that would just shrink the 32 years by maybe a decade. How do others feel about this? 10 years in my opinion is not reasonable. 20 years would be a better estimate. We are probably looking at 20 - 30 years for a desktop player of this strength. And I assume that the MCTS will continue to be refined and improved. Another factor is that Kim could easily be off by a stone or two in either direction - but since he won 2 fast games I would guess that once he got used to playing Mogo he could win consistently with 8 stones - but this is only my guess of course. My estimate may sound pessimistic to some, but this same wild exuberance happened in chess with the famous Levy bet. 10 years later Levy beat the computer winning the bet and 10 more years later he won again. And Levy was not a Grandmaster, he was an international master. - Don I guess I should also go on record as believing that if it really does take 32 years, we *will* have general-purpose AI before then. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] mogo beats pro!
your calculation is for mogo to beat kim, according to kim and the mogo team's estimates. i think that a better thing to measure would be for a computer program to be able to regularly beat amateurs of any rank without handicap. i.e. to effectively be at the pro level. for one thing, this is easier to measure, and for another, it relies much less on mogo staying the same, kim being correct, or some other professional being much better against computer players, for instance. it just requires some machine connected to KGS to be able to attain, say, 9d, and keep it for a month or so. s. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] mogo beats pro!
Congratulations to Mogo team! Twenty years from now, in ``a computer go history'' August 7th 2008: First victory of computer against pro with 9 handicap. By the way, the surge in strength with the 800 processors with respect to the quadcore (old) MogoBot, seemed relatively low, when comparing to the gain with more time (*6 instead of *800). Of course you do not have as many playouts, and the algorithm gets different. But I am curious to know whether you have made studies like self-play of a (smaller) cluster against a single node Mogo with more time (without pondering), to see how more nodes scales against more time? Et encore félicitations ! Jonas ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] mogo beats pro!
- Original Message I still have this theory that when the level of the program is in the high-dan reaches, it can take proper advantage of an opening book. Alas, it may be a few years before enough processoring power is routinely available to test this hypothesis. I know that we duffers can always ruin a perfectly good joseki just as soon as we leave the memorized sequence. Steve: why would this be the case? and where would the book come from? my thinking is that unless mogo created the book itself, playing games like these, against opponents like these, at time controls like this one, then it couldn't possibly be helpful. and even then it might not be helpful. As far as we could see, Mogo was essentially re-creating book knowledge the hard way - using millions of playouts times many seconds to do so. The opening is the same in every game: you start with an empty board or a given number of handicap stones; why spend minutes figuring out the best first move, instead of precalculating that information? As for where it would come from, observation of thousands of pro games would reveal what they do in a variety of standard sequences. This information is not useful if the program cannot play at that level -- lower-level players often botch the followup to joseki, or choose the wrong joseki for the given whole-board situation. But a program which uses joseki to guide search could optimize search. You can reliably say that in certain situations, when you play move A, even the strongest pro is very likely to respond with one of a handful of plays; if this knowledge is part of the search strategy, the search is much more efficient. If you choose to play some other move, it needs to be demonstrably better than the standard replies. A more efficient opening would enable more time to be spent on the complex middle-game situations. - Original Message From: Darren Cook [EMAIL PROTECTED] I do have to ask -- if 1.7 million playouts per second and an hour of playing time are required to reach this level, ... Can Olivier give us more details. A few questions that come to mind: how many playouts per *move* was it using in each of the opening, middle game and endgame? Was it using a fuseki book, and how many moves did the game stay in that book? And once it was out of the book was it all UCT (*) search, or were there any joseki libraries, etc.? I'd also be interested to hear how inefficient the cluster was (e.g. 1000 CPUs won't be doing 1000 times the number of playouts, there must be some overhead). Darren *: Sorry, I've forgotten the new term we are supposed to use. ___ 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] mogo beats pro!
Congrats to the MoGo team for getting system time at SARA for a match. The architecture of the power5/power6 system (2007 july a power5 system was installed and that has been updated to power6 now), is based upon having sufficient RAM and high bandwidth to i/o (for each Gflop a specific amount of i/o bandwidth is there). When they opened the machine (when i stand in front of it at SARA), you can see that a power6 is basically 1 big block of metal that eats a shitlot of power. A single rack with a few powerblocks eats more power than the entire storage of it (which is a huge wall of harddrives). Huijgens has 8 dual core processors, 16 cores a node in short, which are very bad for integers. For most go/chess/checkers type applications a 4.7Ghz power6 single core performs like a 2.4Ghz core2. That is because it focuses upon a high number of instructions a cycle, basically floating point (double precision). The power processor is not so great for most programs for integers. Its IPC (instructions per cycle) is far under 1.0 for most programs. For such type of hardware i therefore did do a PGO run usually with Diep that lasts for 24 hours. After that the compiler software has more datapoints to optimize the software better. That gives a huge speedboost. Each power6 block is connected with a highspeed network to the other nodes. There is 2 physical network lines, i guess one for i/o and one for communication (RAM). So the latency + bandwidth from 1 node to another is real bad compared to the huge crunching power of 1 node, from RAM communication viewpoint seen. A blocked get will be between 5 and 10 microseconds. That's not so fast. From calculation viewpoint seen (instructions a cycle executed), the power6 is not so fast when you need more crunching power than 1 block. So for example for my chessprogram having 1 node would be interesting as that's a lot faster than a PC. I wouldn't ask for more than 1 node though; hashtable is too important for a program like Diep. Even though its nps is low compared to other chessprograms (it has the biggest evaluation function of all chessprograms) it still is doing a 200k nps a core * 16 cores = 3.2 million nps. That's 3.2 million reads and 3.2 million writes to/from a shared global hashtable. Add to that that the network itself then also will deliver 6.4 million transactions a second from the other nodes. So that's a grand total of 12.8 million. Suppose we say this is allowed to eat 10% of our system time, then we would require a latency equal to handling fulltime 12.8 * 10 = 128 million transactions a second or under 10 nanoseconds. In reality however the network is factor 1000 slower in latency, demonstrating the latency problem to its full extend when one uses more than 1 node for game tree search. When using many processors (hundreds) using a global shared hashtable is extra important, as it prematurely terminates by means of a hashtable cutoff millions of searches that are not needed to get done. The necessity of a global shared hashtable was already noticed clearly by Rainer Feldmann long time ago. When not using a shared hashtable last 10 plies at 460 processors, the search depth loss is about 6 to 7 ply at tournament level time controls for Diep. The power6 is a magnicifent machine for software that can use big RAM. Power6 as in SARA has hundreds of gigabytes of RAM. If your software can work on multiple of those nodes very well, one should consider buying for a $10k a big Tesla setup. It has some thousands of cores clocked to 1.x Ghz. You can get one to develop at already for just 520 euro. The number of instructions a second Tesla can execute is a lot higher than power6 can do for the sequential game tree searching codes. Tesla also gets handsdown an IPC of 1.0, higher than Power6. Let's use the IPC = 1.4 at core2 2.4Ghz, that's giving power6 roughly ipc=0.75 Those 800 processors executing : 800 * 0.75 * 4.7Ghz = 2.8T instructions per second. that's quite a lot compared to a simple PC, which of course is a zillion times more efficient thanks to global shared hashtables. A single new card from Nvidia when it is in tesla form is: 240 cores * 1.2 Ghz = 0.36T instructions a cycle. The tesla setups combine a bunch of those cards, giving a similar amount of instructions a cycle. Assuming of course one isn't using double precision floating point for the software. a new nvidia GPU has 60 processors onboard that can do double precision, not 240. I'd estimate the search efficiency of mogo at the same efficiency level like Deep blue. Deep Blue wasn't 11 teraflop. Not even close. Its hardware was of course much faster than that from instructions per cycle seen. It was having an average of 130 million nodes per second. Now for simplistico programs nowadays, A single node eats roughly 1000 cycles at core2. IPC = 1.4 1000 cycles a node 130 mln nps.
Re: [computer-go] mogo beats pro!
On Aug 9, 2008, at 4:16 AM, terry mcintyre wrote: - Original Message I still have this theory that when the level of the program is in the high-dan reaches, it can take proper advantage of an opening book. Alas, it may be a few years before enough processoring power is routinely available to test this hypothesis. I know that we duffers can always ruin a perfectly good joseki just as soon as we leave the memorized sequence. Steve: why would this be the case? and where would the book come from? A thousand years of Go experience? There are many good books on fuseki and joseki. The challenge is encoding that knowledge flexibly and using the information appropriately. Compared to earlier programs, this is one area where the MC programs have taken a step backwards (or rather, a step toward the center). my thinking is that unless mogo created the book itself, playing games like these, against opponents like these, at time controls like this one, then it couldn't possibly be helpful. and even then it might not be helpful. There is an obvious need for adding features to Go programs to make them play more like humans. The public won't buy programs that play too strangely, even if they are objectively stronger. The challenges to MC programs are to: 1. Play more normal looking fuseki. 2. Play joseki moves when available, and use appropriate joseki for the current board situation. 3. Correct seki detection and evaluation. 4. Some sort of sliding komi so programs still play reasonably when far ahead or far behind. 4. Toward the endgame, switch to greedier evaluations that maximize points. 6. Pass instead of filling in territory when all dame are filled. As far as we could see, Mogo was essentially re-creating book knowledge the hard way - using millions of playouts times many seconds to do so. The opening is the same in every game: you start with an empty board or a given number of handicap stones; why spend minutes figuring out the best first move, instead of precalculating that information? As for where it would come from, observation of thousands of pro games would reveal what they do in a variety of standard sequences. This information is not useful if the program cannot play at that level -- lower-level players often botch the followup to joseki, or choose the wrong joseki for the given whole- board situation. But a program which uses joseki to guide search could optimize search. There have already been programs that have used pro game databases for opening moves. Howard Landman's Poka springs to mind. You can reliably say that in certain situations, when you play move A, even the strongest pro is very likely to respond with one of a handful of plays; if this knowledge is part of the search strategy, the search is much more efficient. If you choose to play some other move, it needs to be demonstrably better than the standard replies. A more efficient opening would enable more time to be spent on the complex middle-game situations. Indeed. That is especially beneficial for scalable search algorithms. For example, Orego had a simplistic yet effective fuseki: try to play on all the star points for the first nine moves. That saved quite a lot of time and still obtained a reasonable looking opening. Ian ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] mogo beats pro!
Chris may be right with his implication that I talk too much these days, but just to keep things honest, the quote below is not exactly what I said. I said that others were wondering how much time it will be before the programs are beating the pros. My thought was that programs have advanced 7 to 9 stones in the last few years, and after this match, for the first time I think that programs will likely be competing evenly with pros within a decade. I am shocked to be thinking this ... I certainly did not think this yesterday. Cheers, David On 7, Aug 2008, at 8:47 PM, terry mcintyre wrote: This is from the AGA newsletter: COMPUTER BEATS PRO AT U.S. GO CONGRESS: In a historic achievement, the MoGo computer program defeated Myungwan Kim 8P (l) Thursday afternoon by 1.5 points in a 9-stone game billed as “Humanity’s Last Stand?” “It played really well,” said Kim, who estimated MoGo’s current strength at “two or maybe three dan,” though he noted that the program – which used 800 processors, at 4.7 Ghz, 15 Teraflops on a borrowed European supercomputer – “made some 5-dan moves,” like those in the lower right-hand corner, where Moyogo took advantage of a mistake by Kim to get an early lead. “I can’t tell you how amazing this is,” David Doshay -- the SlugGo programmer who suggested the match -- told the E-Journal after the game. “I’m shocked at the result. I really didn’t expect the computer to win in a one-hour game.” Kim easily won two blitz games with 9 stones and 11 stones and minutes and lost one with 12 stones and 15 minutes by 3.5 points. The games were played live at the U.S. Go Congress, with over 500 watching online on KGS. “I think there’s no chance on nine stones,” Kim told the EJ after the game. “It would even be difficult with eight stones. MoGo played really well; after getting a lead, every time I played aggressively, it just played safely, even when it meant sacrificing some stones. It didn’t try to maximize the win and just played the most sure way to win. It’s like a machine.” The game generated a lot of interest and discussion about the game’s tactics and philosophical implications. “Congratulations on making history today,” game organizer Peter Drake told both Kim and Olivier Teytaud, one of MoGo’s programmers, who participated ina brief online chat after the game. At a rare loss for words in a brief interview with the EJ after the game, Doshay wondered “How much time do we have left? We’ve improved nine stones in just a year and I suspect the next nine will fall quickly now.” - reported by Chris Garlock, photo by Brian Allen Terry McIntyre [EMAIL PROTECTED] “Wherever is found what is called a paternal government, there is found state education. It has been discovered that the best way to insure implicit obedience is to commence tyranny in the nursery.” Benjamin Disraeli, Speech in the House of Commons [June 15, 1874] ___ 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] mogo beats pro!
I still have this theory that when the level of the program is in the high-dan reaches, it can take proper advantage of an opening book. Alas, it may be a few years before enough processoring power is routinely available to test this hypothesis. I know that we duffers can always ruin a perfectly good joseki just as soon as we leave the memorized sequence. why would this be the case? and where would the book come from? my thinking is that unless mogo created the book itself, playing games like these, against opponents like these, at time controls like this one, then it couldn't possibly be helpful. and even then it might not be helpful. s. - Original Message From: Darren Cook [EMAIL PROTECTED] I do have to ask -- if 1.7 million playouts per second and an hour of playing time are required to reach this level, ... Can Olivier give us more details. A few questions that come to mind: how many playouts per *move* was it using in each of the opening, middle game and endgame? Was it using a fuseki book, and how many moves did the game stay in that book? And once it was out of the book was it all UCT (*) search, or were there any joseki libraries, etc.? I'd also be interested to hear how inefficient the cluster was (e.g. 1000 CPUs won't be doing 1000 times the number of playouts, there must be some overhead). Darren *: Sorry, I've forgotten the new term we are supposed to use. ___ 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] mogo beats pro!
First of all, congratulations to the MoGo team. As some have remarked already, the difference in level between the fast games and the slow games was considerable. I didn't think the level of the fast games was anything to boast about. And my opinion is more informed than many other observers who don't understand why MoGo plays so 'stupidly' when behind. The remarks from the kibitzers weren't very flattering. Where I do differ in opinion from most is the remarks from the pro. He played too fast and made a few terrible mistakes at crucial points. He said that MoGo winning the lower-right corner was 5-dan level play but I strongly disagree. It was good play, probably dan- level, but the kicker was that the mistake by the pro was also almost sub dan level. It's a very standard situation that should be very easy for a pro to get right. So I disagree when he says he wouldn't have played stronger with more time, this was a very big mistake that was easy to take advantage of. Then the upper-right corner. I think MoGo played very well there. I don't have the game-record to verify, but I had the feeling the pro could have done a lot better there too with a little more thinking. But I must admit that MoGo took advantage of it very well and I thought the play there was probably 5-dan level or more. To me that was where MoGo won the game. Although there was a point where I thought it might still lose. This was just one game of course but it's an encouraging result. Mark ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] Mogo beats pro: the hardware
Dear all, The machine that was used by MoGo yesterday is the Dutch supercomputer Huygens, situated in Amsterdam. Huygens was provided by SARA (www.sara.nl) and NCF(http://www.nwo.nl/nwohome.nsf/pages/ACPP_4X6R5C_Eng). Huygens was upgraded on August 1 to 60 Teraflops (Peak), so porting MoGo with this short notice for the match was a lot of hard work and stress. But the result showed it was worth it! Huygens is constituted of 104 nodes of 16 dual-cores POWER6 processors at 4.7 GHz each (with 128G of RAM). MoGo was using 25 nodes, i.e., 800 cores and nearly 15 Teraflops. By comparison, Deep Blue was using only 11 Gigaflops. The structure of Huygens with powerful processors and numerous cores per node is ideal for MoGo. It would be less efficient to use a supercomputer with few cores per node, e.g., Blue Gene. The parallelization was performed using Pthreads and OpenMPI. On each node, two search-trees were built, each one using 32 threads. Thanks to the SMT technology, it is actually more efficient to use two threads per core. Indeed, while one thread is looking in memory, an other thread can use the core. Every 350 milliseconds, the nodes were communicating their tree to one an other using OpenMPI. Finally, we would like to help all the people who helped us to port the code, both from the french INRIA/CNRS and the dutch NCF/SARA. Cheers, and see you in Beijing! The Mogo Team: http://www.lri.fr/~teytaud/mogo.html PS: A nice picture of Huygens, and further information, can be found here: www.sara.nl winmail.dat___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] mogo beats pro!
Wow! I've been radio silent for a long time now working on other things some years now, but watching the successes of the new approaches. What incredible validation them... Fantastic! Jeffrey Greenberg www.jeffrey-greenberg.com -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of terry mcintyre Sent: Thursday, August 07, 2008 8:48 PM To: computer go Subject: [computer-go] mogo beats pro! This is from the AGA newsletter: COMPUTER BEATS PRO AT U.S. GO CONGRESS: In a historic achievement, the MoGo computer program defeated Myungwan Kim 8P (l) Thursday afternoon by 1.5 points in a 9-stone game billed as Humanity's Last Stand? It played really well, said Kim, who estimated MoGo's current strength at two or maybe three dan, though he noted that the program - which used 800 processors, at 4.7 Ghz, 15 Teraflops on a borrowed European supercomputer - made some 5-dan moves, like those in the lower right-hand corner, where Moyogo took advantage of a mistake by Kim to get an early lead. I can't tell you how amazing this is, David Doshay -- the SlugGo programmer who suggested the match -- told the E-Journal after the game. I'm shocked at the result. I really didn't expect the computer to win in a one-hour game. Kim easily won two blitz games with 9 stones and 11 stones and minutes and lost one with 12 stones and 15 minutes by 3.5 points. The games were played live at the U.S. Go Congress, with over 500 watching online on KGS. I think there's no chance on nine stones, Kim told the EJ after the game. It would even be difficult with eight stones. MoGo played really well; after getting a lead, every time I played aggressively, it just played safely, even when it meant sacrificing some stones. It didn't try to maximize the win and just played the most sure way to win. It's like a machine. The game generated a lot of interest and discussion about the game's tactics and philosophical implications. Congratulations on making history today, game organizer Peter Drake told both Kim and Olivier Teytaud, one of MoGo's programmers, who participated ina brief online chat after the game. At a rare loss for words in a brief interview with the EJ after the game, Doshay wondered How much time do we have left? We've improved nine stones in just a year and I suspect the next nine will fall quickly now. - reported by Chris Garlock, photo by Brian Allen Terry McIntyre [EMAIL PROTECTED] Wherever is found what is called a paternal government, there is found state education. It has been discovered that the best way to insure implicit obedience is to commence tyranny in the nursery. Benjamin Disraeli, Speech in the House of Commons [June 15, 1874] ___ 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] mogo beats pro!
Thanks for posting the game Eric. When I look back at it it's obvious to me S1 was much better. After the likely sequence of R1, T3, T2, T4, S7, Q1, R7 Black still has a serious weakness at N4. I also still question W's play in the upper-right. I doubt W S15 was a good move and think S19 would have given a much better result. Also O15 I find questionable, better not play here at all and let Black play S17. That way you get the same result as in the game, with the difference that Black added a stone at S17. But I must say that B R16 was a really strong move to make things complicated for W. Hats off to MoGo. Mark On 8-aug-08, at 11:29, Eric Boesch wrote: MyungWan-MoGoTiTan-4.sgf ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] mogo beats pro!
On 8, Aug 2008, at 7:29 AM, Eric Boesch wrote: On Fri, Aug 8, 2008 at 8:04 AM, Mark Boon [EMAIL PROTECTED] wrote: First of all, congratulations to the MoGo team. Ditto! Absolutely an amazing achievement! Where I do differ in opinion from most is the remarks from the pro. He played too fast and made a few terrible mistakes at crucial points. He said that MoGo winning the lower-right corner was 5-dan level play but I strongly disagree. It was good play, probably dan-level, but the kicker was that the mistake by the pro was also almost sub dan level. If it was an outright blunder then it was definitely sub-dan level. I just don't know if that's the case. In discussion one person said So it was an overplay on your part? His answer was With 9 stones I must make overplays. From the nature of what was said after that it is clear that Mr Kim knew how to live in the corner, but chose a variation that he did not think the computer would answer so well that would prevent him from getting sealed into the corner. Maybe (as an alternative to the misclick theory, which would be my other top candidate) Myung Wan I mentioned the misclick theory in the chat and he was emphatic that there was no misclick. deliberately tested Mogo for a blunder He showed us the kind of responses he expected, which he said were 1-Dan level. after Mogo played a very nice squeeze. Myung Wan was disrespecting his opponent to even try to see if r1 would work (but maybe he wanted to see just in case Mogo was that dumb, and he wanted to find out early in the game so he would know how many points he needed to make elsewhere), but s1 is not great either because if Myung saves his bottom chain, then as in the variation starting with move 52 in the attached SGF, Myung is stuck in the corner after black plays s7. It is hard to imagine that the lower right side is worth losing the corner, but maybe the difference is small. To me, r2 looks very good. What do you stronger players think? (I'm only 1 kyu.) MyungWan- MoGoTiTan-4.sgf___ 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] mogo beats pro!
This is from the AGA newsletter: COMPUTER BEATS PRO AT U.S. GO CONGRESS: In a historic achievement, the MoGo computer program defeated Myungwan Kim 8P (l) Thursday afternoon by 1.5 points in a 9-stone game billed as “Humanity’s Last Stand?” “It played really well,” said Kim, who estimated MoGo’s current strength at “two or maybe three dan,” though he noted that the program – which used 800 processors, at 4.7 Ghz, 15 Teraflops on a borrowed European supercomputer – “made some 5-dan moves,” like those in the lower right-hand corner, where Moyogo took advantage of a mistake by Kim to get an early lead. “I can’t tell you how amazing this is,” David Doshay -- the SlugGo programmer who suggested the match -- told the E-Journal after the game. “I’m shocked at the result. I really didn’t expect the computer to win in a one-hour game.” Kim easily won two blitz games with 9 stones and 11 stones and minutes and lost one with 12 stones and 15 minutes by 3.5 points. The games were played live at the U.S. Go Congress, with over 500 watching online on KGS. “I think there’s no chance on nine stones,” Kim told the EJ after the game. “It would even be difficult with eight stones. MoGo played really well; after getting a lead, every time I played aggressively, it just played safely, even when it meant sacrificing some stones. It didn’t try to maximize the win and just played the most sure way to win. It’s like a machine.” The game generated a lot of interest and discussion about the game’s tactics and philosophical implications. “Congratulations on making history today,” game organizer Peter Drake told both Kim and Olivier Teytaud, one of MoGo’s programmers, who participated ina brief online chat after the game. At a rare loss for words in a brief interview with the EJ after the game, Doshay wondered “How much time do we have left? We’ve improved nine stones in just a year and I suspect the next nine will fall quickly now.” - reported by Chris Garlock, photo by Brian Allen Terry McIntyre [EMAIL PROTECTED] “Wherever is found what is called a paternal government, there is found state education. It has been discovered that the best way to insure implicit obedience is to commence tyranny in the nursery.” Benjamin Disraeli, Speech in the House of Commons [June 15, 1874] ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] mogo beats pro!
I enjoyed watching this game. Having trouble with KGS at the moment, or I'd send a game record. Having more time makes a very marked improvement in the quality of play, to a degree which surprised me. The first two games, at 10 and something between 10 and 15 minutes ( Mogo thought it only had 10 minutes until it was restarted ), Mogo made mistakes which kyu players would avoid. The 15-minute game was much better. The 60 minute game, as the pro said, was 2-3 dan level, with a remarkable sequence in the bottom right corner which he says is 5-dan level. I think computer clusters will beat pros within a decade; combination of better algorithms and cheaper processor power will ensure that. Many thanks to everyone who made this possible, especially to the team of Mogo and whoever donated the European Supercomputer. I do have to ask -- if 1.7 million playouts per second are required and an hour of playing time are required to reach this level, is it possible to greatly improve the efficiency? Humans surely don't process that much information to accomplish a very high level of performance. The Mogo team has done heroic work. I understand they are working on adaptive playouts which would better utilize information about the board, and presumably gather higher quality information with less effort. All the best to you wonderful people for making this program possible! Terry McIntyre [EMAIL PROTECTED] ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] mogo beats pro!
We had a bit of discussion w/ the Mogo team after the match. ( I am writing this from the US Go Congress in Portlland, OR ), and Olivier said that Mogo no longer uses a book; it was found to be ineffective in their research. I am wondering, of course, if a book would be more effective now that Mogo has such impressive power - but unfortunately, the supercomputer was only lent for this one match. They can't easily test the hypothesis that a strong program on a massive cluster could make effective use of a good opening book. Mogo no longer uses UCT. Terry McIntyre [EMAIL PROTECTED] - Original Message From: Darren Cook [EMAIL PROTECTED] To: computer-go computer-go@computer-go.org Sent: Thursday, August 7, 2008 9:24:00 PM Subject: Re: [computer-go] mogo beats pro! Great news. Well done to the Mogo team. John, if I can just find 3000 CPUs lying around I might actually win our bet ;-). I do have to ask -- if 1.7 million playouts per second are required and an hour of playing time are required to reach this level, ... Can Olivier give us more details. A few questions that come to mind: how many playouts per *move* was it using in each of the opening, middle game and endgame? Was it using a fuseki book, and how many moves did the game stay in that book? And once it was out of the book was it all UCT (*) search, or were there any joseki libraries, etc.? I'd also be interested to hear how inefficient the cluster was (e.g. 1000 CPUs won't be doing 1000 times the number of playouts, there must be some overhead). Darren *: Sorry, I've forgotten the new term we are supposed to use. -- Darren Cook, Software Researcher/Developer http://dcook.org/mlsn/ (English-Japanese-German-Chinese-Arabic open source dictionary/semantic network) http://dcook.org/work/ (About me and my work) http://darrendev.blogspot.com/ (blog on php, flash, i18n, linux, ...) ___ 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] mogo beats pro!
To answer one other question: we were told that Mogo scales linearly. The supercomputer has a very high-bandwidth interconnect. The Mogo team was unable to release more architectural details at this time. To reiterate on another question, from what the team said, no book, no joseki, just raw search using billions and billions of galloping CPU cycles. Some of the plays were described by the pro as 5 dan level -- effectively, Mogo generated joseki from whole cloth. I was impressed. Check out the KGS records. If my memory is correct, the userid was MogoTitan. I'd love to hear feedback from stronger players, but it seemed to me that, as Mogo was given more time, its opening and middlegame play was markedly better. I have a question: Is the allocation of time front-weighted, to take advantage of the fact that much less effort is required to calculate end-game plays, since the playouts are so much shorter? The playouts are shorter, the search tree has fewer branches; the time needed should decay rapidly. I still have this theory that when the level of the program is in the high-dan reaches, it can take proper advantage of an opening book. Alas, it may be a few years before enough processoring power is routinely available to test this hypothesis. I know that we duffers can always ruin a perfectly good joseki just as soon as we leave the memorized sequence. - Original Message From: Darren Cook [EMAIL PROTECTED] I do have to ask -- if 1.7 million playouts per second and an hour of playing time are required to reach this level, ... Can Olivier give us more details. A few questions that come to mind: how many playouts per *move* was it using in each of the opening, middle game and endgame? Was it using a fuseki book, and how many moves did the game stay in that book? And once it was out of the book was it all UCT (*) search, or were there any joseki libraries, etc.? I'd also be interested to hear how inefficient the cluster was (e.g. 1000 CPUs won't be doing 1000 times the number of playouts, there must be some overhead). Darren *: Sorry, I've forgotten the new term we are supposed to use. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] mogo beats pro!
Well done, Mogo team ! terry mcintyre wrote: moves,” like those in the lower right-hand corner, where Moyogo took Typo :-) Rémi ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] mogo beats pro!
... no book, no joseki,...Mogo generated joseki from whole cloth. ... seemed to me that, as Mogo was given more time, its opening and middlegame play was markedly better. If it is basically reinventing opening theory from scratch each time then that makes sense. (Though I suppose there is indirectly some go opening knowledge (aka good shape) in the heavy playout algorithms.) Darren -- Darren Cook, Software Researcher/Developer http://dcook.org/mlsn/ (English-Japanese-German-Chinese-Arabic open source dictionary/semantic network) http://dcook.org/work/ (About me and my work) http://darrendev.blogspot.com/ (blog on php, flash, i18n, linux, ...) ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] mogo-pr-4core on cgos 9x9
Dear CGOS watchers, :) Mogo-pr-4core is the public release (not big float) version of MoGo and running on Intel Q9550, latest 45nm chip of Core2 micro architecture with larger L2 cache (2 x 6MB), at 3.4 GHz (8.5 x 400MHz; overclocked from rated 2.83GHz) on ASUS P5K-VM motherboard. Its performance is almost tie against mogo-big-4core but seems better to others. Somewhat interesting. -Hideki -- [EMAIL PROTECTED] (Kato) ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo/professional challenge
on 19x19 board, any thought on getting computer to win all the games with handicap 9(may be started at 13), then improve to handicap 8,7...? Olivier Teytaud wrote: Has the program become that much stronger on 9x9 recently? (Compared to the version was trying?) *Parallelization: MPI == ~80% vs no mpi in 9x9 (for same number of cores). *Monte-Carlo improvement == strongly depends on number of simulations and number of cores (as the multi-core reduces the influence of the computational overhead), ~55% I guess. *Openings: 58%, for games with constant time per move (should be higher for games with given total time), if we only keep the openings which are still efficient in the new version of the code. Human-based openings do not work :-( *less interestingly, we have a better hardware than at that time (more cores, more GHz). == no doubt that this mogo is by far stronger than the one at Amsterdam 07. The improvement is much higher in 19x19, but humans are really too strong in 19x19 :-) ___ computer-go mailing list computer-go at computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ - Be a better friend, newshound, and know-it-all with Yahoo! Mobile. Try it now.___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] MoGo/Taranu challenge in Paris
I have put a report of the weekend's challenge games between MoGo and Catalin Taranu 5p at http://www.computer-go.info/tc/ mainly to make it easier for people to find the game records. Nick -- Nick Wedd[EMAIL PROTECTED] ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo/Taranu challenge in Paris
I have put a report of the weekend's challenge games between MoGo and Catalin Taranu 5p at http://www.computer-go.info/tc/ mainly to make it easier for people to find the game records. Thanks a lot for that. Some points are wrong however, below some informations about the errors. The hardware was provided by Bull and operated by Bull; we have never had an access to the hardware ourselves, in the mogo-team. Only the solution used when the initial hardware was down was operated by me. The hardware used for calibrating comes from Grid5000, but as Grid5000 does not agree for the use of its hardware for Go competitions, it was only for calibrating and experimenting. The hardware in case of trouble, which has been used for two games, is provided by Université Paris-Sud. The code was originally written by Sylvain Gelly and Yizao Wang, with ideas from Remi Munos and minor contributions from me. Rémi Coulom has also helped the initial developments and we had some discussions with him at various steps of the development. Then, the code has been improved by Arpad Rimmel, Jean-Baptiste Hoock, Julien Perez and me. Thomas Herault and Vincent Danjean also helped for profiling various aspects of the parallelization. This is a small list of contributors, as we have opened the cvs access to many people and many people have suggested or coded some points: a complete list of contributors is in http://www.lri.fr/~teytaud/crmogo.html (in french, an english version should come soon). This web page is itself a draft, in progress; I'll include more informations later. For the first game, the game was played by the cluster from Bull during 13 moves, then the cluster was down (hard drive full). I then launched the second machine, a 4-core system. For the second and third game, the game was played by the cluster with no bug; one win for Catalin, one win for Mogo. For side games, played around the challenge against Catalin, there was no bug; mogo won both times. In 19x19, the cluster lost its connection to internet and therefore I launched once again the second solution. The cluster came back later in the game. I am not able to remember at which times (perhaps in the KGS record). In my humble opinion, there was no problem for the clock in 9x9 games; only in 19x19, we were afraid that the KGS time might be different from the physical clock, but the trouble was minor as mogo resigned before any trouble. Thanks, Olivier___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo/Taranu challenge in Paris
The hardware in case of trouble, which has been used for two games, is provided by Université Paris-Sud. Precisely: LRI, Université Paris-Sud.___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo/professional challenge
It was 2 cores 2.6GHz. (intel core2 duo). sorry, I believed it was the tipi. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
[computer-go] MoGo/professional challenge
This Easter weekend, there will be a challenge between MoGo running on a very powerful system, and Catalin Taranu, 5-dan professional. The following is from the info of The Enclave room on KGS. It is confirmed by the page http://paris2008.jeudego.org/ quotation starts A unique challenge will be held in parallel to the Paris Go tournament : Mogo, currently one of the best Go programs in the world, will challenge the professional Go player Catalin Taranu 5P. Mogo has all the computing power of INRIA with hundreds of super-computers in a network. The winner will be chosen at the end of 9x9 games after 3 rounds of 2x30-minute sudden death. A 19x19 exhibition will be held on Sunday. Events are being followed live on KGS! They will be shown by 'iagochall'. Saturday: 3/23/08 3:00 PM Game I (9x9) Game II 9x9 Game III 9x9 Played with 1.5 hours from the start of one round to the next Sunday: 3/24/08 3:00 PM Exhibition game (19x19) Monday: 3/25/08 11:00 PM Debate with participants quotation ends The time zone quoted above is GMT; that in the http://paris2008.jeudego.org/ page is French time, GMT+1. Nick -- Nick Wedd[EMAIL PROTECTED] ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo/professional challenge
For information on the mogo/pro challenge: - during preliminary tests, mogo has won 4/0 against a very high level human; at that time we were just very very very happy :-) - some other humans, supposed to be weaker, have however won some games at that time (before the nakade correction); - the Nakade weakness is currently assumed to be solved, but I'm not sure of that - at least mogo solves the old known nakade situations and is stronger than the old mogo; at that time we were happy again :-) - another improvement is that we currently have access to much hardware than during tests above; - but, a human, supposed to be weaker (non professional level, 5Dan however) has found some trick to win against mogo; this is not the nakade, but this is seemingly stable, and I am just not able of explaining how he can do that; he has shown me situations and says that in this kind of situations, mogo makes an error, but I just don't understand the common point in these situations. If we understand something we will post details here (at least the sgf files)... - in 9x9, the MPI (multi-machine) version of mogo wins with probability 80% against the non-MPI version. The speed-up is better in 19x19 and will be detailed later, after extensive experiments - the focus was on 9x9 until now due to the challenge. - once again, very strong improvements in front of old versions of mogo leads to disappointing improvements against humans. However, I think that the best 9x9 go programs (mogo and others) are currently difficult opponents for high level players. Everything is under writing for publication and will be sent on this mailing list. Some technical details: - due to concurrency in memory access, heavier playouts come for free. If playouts are heavier (computationally more expensive) the speed-up becomes better. The nakade-problem involves heavier playouts, but the computational overhead is almost canceled by the speed-up improvement, as the speed-limit on 8-core machine is due to concurrency in memory access (for modifying the tree) more than to computational cost. - (very) unfortunately, the opening books generated for mogo without nakade are seemingly poor for mogo with nakade... this has destroyed weeks of work. If mogo wins the challenge, I'd like to point out that this is a collective success of the computer-go mailing list - without gnugo, crazystone, cgos, kgs and so on, mogo would just not exist. Thanks to all of you for that. I regret that due to some restrictions, we have not published every detail before, but it was just a matter of weeks and I'm happy that everything will be published soon, and if we loose the challenge I hope someone else will win something similar soon :-) Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo/professional challenge
How well does the nakade improvement perform on 13x13? -- robert jasiek ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo/professional challenge
How well does the nakade improvement perform on 13x13? no idea on 13x13, but it does not work on 19x19 (seemingly, perhaps we just need tuning...). Also, it works only, in terms of success rate against the old mogo, for sufficiently large number of simulations per move. Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo/professional challenge
This event sounds very interesting! Saturday: 3/23/08 3:00 PM Saturday: 3/22/08 3:00 PM is right? Regards, Hiroshi Yamashita ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] MoGo/professional challenge
Saturday: 3/23/08 3:00 PM Saturday: 3/22/08 3:00 PM is right? Hi; it's saturday 22. Olivier (stress++) ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/