At move 162 Many Faces also likes C7, with an 83% win rate. There seems to be a problem specific to Fuego in this position.
David > -----Original Message----- > From: [email protected] [mailto:computer-go- > [email protected]] On Behalf Of [email protected] > Sent: Thursday, November 25, 2010 11:24 PM > To: [email protected] > Subject: Re: [Computer-go] ManyFaces swindled of victory? > > In the position at move 162 Valkyria would simply play C7 with a score of > 80%. > The eval of the position looks very straightforward to Valkyria. > > My guess it that in this position Fuego as number of problems due to > bugs perhaps in many patterns on the board that is specific to Fuego. > Or that it later in the game evaluated some bad moves better than good > moves ending up in positions very many local misevaluation lead to a > global collapse. this is not the same as the semeai weakness. But is > the same in the sense that tree search cannot resolve a position with > many local tactical possibilities that is mishandled by the playouts. > > Best > Magnus > > > Quoting [email protected]: > > > I can only speak from one example from 2009 of me (4d EGF) playing > > against a strong Monte Carlo program (Fuego) on 19x19, Fuego taking > > 3 stones handicap. > > > > It was a game on a turn base site with one day per move thinking > > time. Fuego was playing multiple games on that site and on 19x19 it > > was allowed one hour thinking time per move per game if I remember > > well. I usually don't use more than one or two minutes per move in > > these conditions. > > This is the game: http://www.online-go.com/games/board.php?boardID=176261 > > > > Up to move 107 things were going well for Fuego. I was quite > > impressed be it strength. Then Fuego started a ruthless attack and > > it was very succesful. By move 128 I was in big trouble. I despately > > tried to wriggle my way out of the situation, but by move 161 > > things had turned even worse for me. I felt I should resign, but I > > couldn't bring myself to it. Then Fuego played 162. A totally > > irrelevent move, almost a pass, allowing me to play 163 to save my > > group and regain hope. And from then on Fuego played more > > irrelevant, incromprensible weak moves until its totally won > > position had turned into a lost position. > > > > I don't know if it has to do with semeais or multiple fights or > > simply a bug. This game did not seem to have an exceptionally large > > number of fights or semeais to me. I was surprised by the transition > > from consistently strong play to consistently weak play. From move > > 162 it felt as if a was playing a different, much weaker player. > > > > I know Fuego is not Many Faces, but I read the same issue about > > fights and semeais applying to all MC programs. Yet I felt that > > semeais and fights were not the problem in this particular game. I > > don't know what was. It felt like a general collapse. > > > > Dave > > > > > > ________________________________ > > > > Van: [email protected] namens David Fotland > > Verzonden: za 13-11-2010 17:52 > > Aan: [email protected] > > Onderwerp: Re: [Computer-go] ManyFaces swindled of victory? > > > > > > > > It is certainly true that strong programs today are weaker at semeai > > than people of the same rating. This must mean that the programs > > are stronger in other areas than equal ranked people. This gives me > > hope that when Many Faces plays semeai properly it will get a big > > jump in strength. > > > > > > > > David > > > > > > > > From: [email protected] > > [mailto:[email protected]] On Behalf Of terry mcintyre > > Sent: Saturday, November 13, 2010 8:33 AM > > To: [email protected] > > Subject: Re: [Computer-go] ManyFaces swindled of victory? > > > > > > > > Don, as you say, humans are very good at discerning patterns - and > > the game of Go is all about patterns. > > > > > > > > Now, in some cases, the pattern-matching ability can lead humans > > astray, but in other cases, it's a done deal. > > > > > > > > For example, we hashed over the concept of nakade a while back. > > These are patterns which strong humans recognize at a glance. Groups > > with certain shapes are mathematically, provably, totally, dead > > beyond hope, assuming proper play. At that time, many programs were > > weak in that area. Now, strong programs usually do not fall into > > such simple traps. > > > > > > > > In addition to "dead beyond hope" and "certainly alive", strong > > humans also recognize "can live with ko" and "seki" shapes - again, > > beyond a shadow of a doubt, as mathematically certain as the sunrise > > in the morning. > > > > > > > > The analogy with the stock market misses the mark because the stock > > market has many millions of independent actors ( the human beings > > who make buy and sell decisions ) who act upon many billions of > > facts, most of which are inaccessible to the punditocracy who try to > > make sense of the markets. (would-be pundits are also often > > handicapped by inferior models, but that's another tale.) > > > > > > > > Go is a game of complete information played by exactly two players, > > if we ignore rengo or phantom variants; when the position is > > simplified enough ( something which strong players actively seek ), > > the result is mathematically provable - and well within the limits > > of what humans can do with their pattern recognition facilities. > > > > > > > > The topic here is that of large semeai; many games have shown that > > programs are vulnerable to misjudging the outcome; Darren Cook has > > written up a page describing the problem with random playouts and > > complex semeai which depend upon precise move ordering. In many > > cases, semeai are won or lost by a single play, and one must play A, > > B, C, D precisely in order, in response to any one of Z, Y, W, X, > > etc. A large class of problems depend on "if my group A loses a > > liberty, I must take a liberty from group B in such-and-such order." > > Strong humans read out these problems and play them correctly. > > Strong programs have been observed to fail. > > > > > > > > Many well-known joseki create a local situation with such "win by 1" > > capturing races. Usually, the person who loses the local race gains > > something - "influence" or "thickness" in compensation; strong > > professional players consider these sequences to lead to equal > > results for both players. If a player takes that compensation, > > converts it to cash ( territory ), and also manages to swindle a > > program out of winning the local semeai, the player can easily win; > > it's like being able to write off a $500,000 mortgage and keep the > > house, because the lender made a mistake in the paperwork. > > > > > > > > It used to be fairly easy to set up a ladder and a ladder-breaker, > > and programs would still play out the ladder as if the ladder > > breaker were not present - a huge misconception. Strong programs > > don't seem to fall into that trap anymore - but they do fall for > > semeai which are conceptually similar, in a mathematically provable > > sense. > > > > > > > > While I present a single game to illustrate my case, I generalize > > from many games. It's still merely a black box analysis, however; I > > leave it to the programmers to "open the box" and discover how the > > internals map to the observable externals. > > > > > > > > There is a difference between Go and Chess. In Chess, only one thing > > matters: put the other guy in checkmate, and you win, even if all > > you have on the board is one king and one pawn, and the other player > > has a dozen pieces. In Go, once you achieve a territorial > > advantage, you need only keep what is yours. Going back a year or > > two, programs were not very good at keeping what was theirs; they > > played odd yose moves which yielded up territory without gaining > > anything in terms of improving their winrate. Semeai are the > > midgame equivalent - moves which are mathematically constrained in > > ways which significantly alter the real status of the game, as > > opposed to the hypothetical winrate of any algorithm which does not > > understand those constraints. > > > > > > > > Terry McIntyre <[email protected]> > > > > Unix/Linux Systems Administration > > Taking time to do it right saves having to do it twice. > > > > > > > > > > > > > > > > > > > > > > > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
