I think your question boils down to answering what is meant by "evaluate". Chess has a heuristic that is easy to compute and gives a good evaluation. Go lacks this. While probably an inferior evaluator, the Bouzy 5/21 score estimator is an example from go that can be quite slow. UCT (or generically Monte Carlo) can "evaluate" a position fairly quickly (maybe 1k-100k per second depending on how heavy the playout is), they don't give a reliable estimate. To improve this, they end up reevaluating positions more than once (maybe 100 times?) to get a more reliable estimate.
On Mon, 2008-01-14 at 17:44 -0800, mingwu wrote: > Hi, > > I read on the web, and some other places that most Go programs can > only evaluate "a dozen" of moves per second. Is this still true today > on a typical machine, say, single 2GHz CPU, 2GB memory? > > And if this is still true, how can we make it faster? > > To make the question more precise, I define a board updates as: > suppose the program places a new move on an existing board, and then > update all the blocks, dragons, eyes, connections, territory ... info, > and output the evaluation as a score ( e.g. B leads by 15.5 points). > > I also read that UTC programs choose a move by running lots of > simulations, are their update speed any faster? Or they evaluation > lots of boards, but for each move they only calculate some very simple > information (to me, it will still be a surprise, because to evaluate a > Go board, one at least have to know the life/death of each dragon, > that would require lots of computation, which I think is the primary > reason that why Go programs is slow compared to Chess programs that > can evaluate thousands or even millions of boards per second). > I'm just curious, any info / thoughts / comments is appreciated. > > > _______________________________________________ > computer-go mailing list > [email protected] > http://www.computer-go.org/mailman/listinfo/computer-go/ _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
