On Sun, 2008-10-12 at 11:35 +0200, Denis fidaali wrote: > > Hi. I recall somehow that don ran some experiments with min-max(alpha-beta) > algorithms a while ago. > So i wonder if anyone had hard-data about some "min-ma"x equivalent > algorithms using "winrate", and "light-simulations". > I'd like to know what parameters were used back then, and as much of the > actual experiments results as possible > (Games against gnugo, scaling, games against amaf bot, cgos rating etc.. as > long as it concerns min-max&light-playout) > Thank you in advance.
I tried all kinds of different things and don't have numbers. The basic idea was to use a bunch of MC sims directly as an evaluation function. For instance 1000 simulations. You can use tricks to cut many of these off early when it looks like they will be futile. But I'm sorry I can't report much on them. I found that the best strength/effort ratio seemed to come from reducing the number of play-outs as depth increased. I think such a thing has a chance to succeed if the idea was more fully developed. There are many opportunities for tree pruning and so on that I barely touched on. There is this thing in computer chess called "Late Move Pruning" which I believe could be applied aggressively, but I barely got into this kind of things. MCTS boils down to highly selective searches with very low branching factors. Modern day chess programs also have very low branching factors compared to a few years ago - so I don't believe alpha/beta is completely out of the question if done right. - Don > I have made a few tests with AMAF&min-max in the past few days, all giving > terrible results. My assumption is that my implementations where faulty. > So i have to design a path for getting it right somehow ... Any hard data > would help me a lot there in knowing how close i am to having it right. > _________________________________________________________________ > 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/
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/
