-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
> It still has the scent of alpha-beta-with-some-evaluation-function, > which is probably not the right paradigm for go. No matter how you slice it, the problem is evaluation. I'm a strong believe in tree search, but it must be combined with knowledge in the right proportions. UCT is a very clever tree search algorithm, very smart because it's incredibly selective and yet it uses on the fly experience to simulate GO knowledge. I think it also happens to simulate how we humans play better than even the pure knowledge and pattern approach. Humans use pattern knowledge at a low level and imagination at a high level and good UCT implementations simulate this well in my opinion. In my opinion, UCT is the best thing we have right now. If UCT had not been discovered, then Hsu's approach is perfectly reasonable and would be the best known try although it would surely produce weak play. But I think he is capable of making a piece of hardware using alpha beta that will beat the current best programs. I have done experiments with patterns that suggest that it would be possible to build a heavily pruned tree which could be used in a global searcher using ordinary alpha beta pruning. The amount of pruning possible (in a reasonably fast way) is hard to determine, I never found enough to make a global searcher very feasible in 19x19. The idea is to use patterns to find moves that have very low probability of being playable. For a scalable alpha beta searcher the probability of being playable would have to be considered by depth (prune more near leaf nodes.) So I personally believe a feasible alpha/beta searcher is possible, if a high quality evaluation function is combined with a highly selective global search in a pragmatic way. One has to be very aggressive about finding moves that can be thrown out and this is tough problem. It's not hard to find a few moves that can be pruned, but it's difficult to eliminate the MAJORITY of moves without throwing out the baby with the bathwater. - - Don > > We'll see. > AvK > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ > -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.6 (GNU/Linux) Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org iD8DBQFHAnd2DsOllbwnSikRAoFWAJ46NF6RFmyd0+xKzDirD8ZmCfsA3ACfcntX HubWY0nNQa8bDtJ3JsMLefA= =LwC5 -----END PGP SIGNATURE----- _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
