On Thursday 03 June 2010 13:10:06 terry mcintyre wrote: > At a birds-eye layman's level, the chess move tree has on average 20 > legal moves; the go tree has over 100; it is much bushier.
The average is closer to 36 for chess, but anyway, the size of the search tree caused by the raw (not the effective) branching factor is quite irrelevant. What matters is how much you can prune it with a reasonable degree of confidence, and my experience so far seems to indicate the real difference here isn't that big. It seems to be much easier to statically determine which moves are good and bad in go than in chess. (The opposite of the positional evaluation!) At least in my case, my Go program searches deeper for the same amount of nodes compared to my chess program. What's more: this isn't actually dependent on the size of the board! I do not believe that the branching factor of Go is actually that much of an issue for the tree search. The problem is the evaluation. The breakthrough in the last few years was the realization that Monte Carlo makes a "good enough" evaluator and that it can be incorporated in a tree search. > Chess offers a > fairly simple evaluation: if you are down on material, you are either > dead lost, or you sacrificed material to obtain a mate-in-n moves. This is quite simply not true at any reasonable level of play. > With Go, evaluation is not so simple; you need to be able to determine > the life-and-death status of numerous groups, the value of ko threats, > and other fairly subtle factors. No difference with chess here. -- GCP _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
