ivan dubois wrote: > I think there is some missconception about this "infinite scalability of MC" > stuff. > There is no misconception. This is clearly a PRACTICAL concept.
> Theory is only usefull when the model it is based on shares important aspects > with reality. > Which it does clearly. > In theory, 19*19 Go is a finite game, wich can be analysed in a finite amount > of time. So ANY algorithm that solves the game at some point is, in theory, > infinitely scalable. For example, the folowing algorithm is infinitely > scalable : > Analyse the complete mini-max tree of the game. If enough time to finish, > returns the correct move, if not, return a random move. > > Now, is this algorithm really scalable, in practive ? Of course not. > This is not a scalable algorithm, it only solves the game and it takes too long for that. Put an evaluation function and iterative deepening and it's an example of an algorithm that scales, but very poorly. (Unless it's really implemented well with alpha/beta, progressive selectivity, excellent evaluation, etc.) > I support the idea that MC is infinitely scalable (in the reality) ONLY if > the play-out part does not have severe misconceptions. That's not the definition of infinite scalability. You seem to believe that infinite scalability is suppose to mean that it plays perfectly with a few seconds of thinking time. Infinite scalability simply means you can add a modest amount of time and expect to see a tangible (we are talking reality here) improvement. Nobody seriously expects perfect play with a few minutes of thinking and you are way off base if you thought that. > So i think that currently, only MC based on uniform playouts is infinitely > scalable. > I don't know what you mean by that, but heavy play-outs will ALWAYS outperform uniform random play-outs and any number of play-outs. Unless the heavy play-outs are actually horrible wrong, but I'm talking about anything proven to be superior at low levels such as the heavy play-outs of every MC program today. > It is well know that even Mogo has troubles with big eyes (he thinks a big > eye gives life, wich is false). This problem can not be solved with more > computing power (excep absurdly big, of course you can always mini-max to the > end). > > ----- Message d'origine ---- > De : Alain Baeckeroot <[EMAIL PROTECTED]> > À : computer-go <computer-go@computer-go.org> > Envoyé le : Mardi, 22 Janvier 2008, 22h13mn 26s > Objet : Re: [computer-go] Is MC-UCT really scalable ... is a troll > > Le mardi 22 janvier 2008, David Fotland a écrit : > >> 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. >> >> > I think Remi posted a game of CrazyStone on 19x19 commented by one pro > who said "this move is 2 dan level". > So far no go program got such analyse, and the also the huge novelty > provided by MC/UCT programs is they have a _real_ strenght with their > own style, like humans: > play solid when winning, and play hamete (tricky moves) when losing. > > On kgs MC programs played hundreds (if not thousands) games against humans, > and no doubt they fully merit their rank, and no doubt they are improving. > > Infinite scalability is a theoricaly proved property, so please > don't feed the troll :-) > > Alain > > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > 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 > 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/
