Hello Don, I think you are mostly right, but there is something you seem not to realise. MC players are very good at "scoring". They know when they will win for sure by 0.5. However, human players, even strong players, generaly never have such accurate counting skills. So if they are ahead by 0.5, there is no way they can anticipate it at as a sure win, because it would require amazing couting skills. Actually, a lot of strong players never count during the course of a game, especialy during fast games. So often, the most rational way of playing is to be conservative with the points you have, and just always play the natural, conservative defending moves. It may be too hard for human to reason about "probability of winning", it is easier to reason about points you have or not. Of course, i think the ultra rational way based on accurate couting by MC programs is a great strength they have. Its a specific strength of MC that human, unfortunately for them, do not have.
----- Message d'origine ---- De : Don Dailey <[EMAIL PROTECTED]> À : computer-go <[email protected]> Envoyé le : Mardi, 22 Janvier 2008, 21h47mn 39s Objet : Re: [computer-go] Is MC-UCT really scalable against humans? David Fotland wrote: > I didn't say that :) Please read what I wrote. > No, I was thinking ahead, not quoting you. I was just covering my bases, anticipating what I thought my be a likely response (and not necessarily from you.) > The UCT programs often find moves that are unconventional. This makes > patterns that aren't in the database, so the traditional programs can't > cope. People are a little more flexible, especially strong players, and can > still find good responses to unconventional moves. > > I don't know what you mean by "unsound". That's chess terminology and I > think it refers to a trick move that should lead to a loss against good > play. Unsound just means a move that is really wrong. It could be tricky to find the right move or might just be an outright bad move. > In go we talk instead about moves that lose points. UCT programs > often play moves that clearly lose points in an attempt to maximize the > probability of winning. These moves are unconventional because people learn > the optimal moves in standard situations. Weaker players and traditional > programs can have trouble finding the best response to these non-optimal > moves. There is an isssue whether these are really non-optimal or not. To me an optimal move preserves the win if a win is possible and nothing else matters. I think it's elegant to not play greedy when another moves is better in the game-theoretic sense. > So the UCT programs gain strength for two reasons, first, because > the maximize the probability of winning rather than maximize the score. And > second because this causes them to play unconventional moves, which make the > game more difficult for their opponent. They don't do this. By the time they start playing unconventional moves they have already won or lost the game. I think this is a huge misconception. It's true that they may fail to defend a big group and attack a small group because it is slightly more likely to win - but that's a pretty rational thing to do - going for the big group may be more human-like but less logical so I don't care if that confuses a weaker opponent. (I suppose the blunder the weaker player would make is to defend the big group and not give due resistance to defending the small group - but for sure there is no good reason to program a program to make the irrational decision because it makes weak players more comfortable.) > There is nothing inherently wrong > with playing unconventional moves that are still good. Strong players do it > whenever they can to make the game more difficult. > I would imagine that the stronger you are, the more you would be concerned with playing for the win instead of points. > David > > >> Please don't say the style is to "find an unsound move that >> is >> difficult to defend", that's not what it's trying to do, it's just >> trying to find a move that it is IMPOSSIBLE to defend, and if it >> >> - Don >> >> >> >> David Fotland wrote: >> >>> I share this opinion. 9x9 was a good simple test to get things >>> >> started, but >> >>> go is a 19x19 game. 9x9 has limited interest. An analogy for chess >>> programmers would be if a group of people worked on programs to solve >>> >> rook >> >>> and pawn endgames. This kind of chess endgame is a good test to try >>> >> out >> >>> algorithms, but if they claim to be making strong chess programs, at >>> >> some >> >>> point they have to implement the full game. In go it turned out that >>> >> to be >> >>> good at 19x19, some new algorithms were needed (patterns and heavy >>> playouts). I think that to take the next step in 19x19 strength the >>> programs will need to be stronger at life and death. >>> >>> >>> >>> The UCT-MC programs do particularly well against traditional programs >>> because they expose the brittleness inherent in the pattern databases >>> >> they >> >>> use. Strong humans are not so easily beaten by playing >>> >> unconventional and >> >>> somewhat inferior moves. >>> >>> >>> >>> David >>> >>> >>> >>> Personally I think 9x9 is good for trying out ideas. But in the end, >>> >> if it >> >>> doesn't play well on 19x19 then I don't care one bit how well it >>> >> plays on >> >>> 9x9. Just my opinion of course. >>> >>> >>> >>> Mark >>> >>> >>> >>> >>> >>> >>> >>> --------------------------------------------------------------------- >>> >> --- >> >>> _______________________________________________ >>> computer-go mailing list >>> [email protected] >>> http://www.computer-go.org/mailman/listinfo/computer-go/ >>> >> _______________________________________________ >> computer-go mailing list >> [email protected] >> http://www.computer-go.org/mailman/listinfo/computer-go/ >> > > _______________________________________________ > computer-go mailing list > [email protected] > http://www.computer-go.org/mailman/listinfo/computer-go/ > > _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/ _____________________________________________________________________________ Ne gardez plus qu'une seule adresse mail ! Copiez vos mails vers Yahoo! Mail http://mail.yahoo.fr _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
