On Thu, Aug 14, 2008 at 12:00 AM, Gian-Carlo Pascutto <[EMAIL PROTECTED]> wrote: > Mark Boon wrote: > >> Not an expert on AB-search or UCT search but there's a subtle >> difference I think. In AB search, if some processors have been >> searching in a branch that is subsequently cut off, the work is 100% >> wasted. In UCT there's no such black-and-white cutting. If you do >> sampling in what then turns out not to be the optimal branch, further >> sampling in the optimal branch may cause the UCT-value to drop to a >> point where you'd otherwise start sampling the sub-optimal branch >> again. So in that case you gain back part of the 'wasted' work. >> >> I'm also inclined to think that the 'chunks' of work that are wasted >> tend to be larger for AB-search than for UCT. But I'm not 100% sure. > > The problem is that the optimal settings for UCT appear to be much stronger > on the exploitation side than on the exploration side, making it much more > likely that such work is really wasted. > > It is not so obvious to me any more what the differences are between a state > of the art UCT searcher (i.e. with nearly no exploration) and a state of the > art alpha-beta searcher (i.e. with heavy late move reductions).
One might consider heuristics like AMAF, pattern knowledge, etc. to be simply a more effective way to guide exploration. The UCB term has no domain-specific knowledge. It works surprisingly well but it should be no surprise that one can do better with domain-specific knowledge. > They both just hammer out the mainline deeper until the score drops a bit, > or an alternative rises strongly. > > UCT without exploration is just minimax, no? The move/path selection behaves like a best-first minimax, but where are the min/max operators in the updates? Erik _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/