On Mon, 2008-08-11 at 20:10 -0700, David Fotland wrote: > Sorry, but I can’t let this statement go past. The go programs in the > 90s did local search, but not much global search. For example Many > Faces did a one ply global search, with a variable depth quiescence > search. I added an alpha-beta search to Many Faces last year, and it > made a huge improvement in strength. So it is not true that > alpha-beta pruning hit a roadblock. >
I never doubted alpha-beta but when you say alpha-beta and GO in the same sentence, people automatically believe the program is going from 99% evaluation to 1% evaluation and 99% stupid. In fact you are still spending most of your time evaluating positions. I'm still not convinced that you can't make a strong alpha beta GO program if you have some imagination. It cannot just be a converted chess program, it has to be different, but still alpha beta at heart. It would have to be extremely selective. - Don > > > For me, the big advantage of UCT/MC is that it eliminates the huge, > slow, buggy evaluation function. Simple playouts are much much easier > to make bug free. Bugs in the evaluation function caused many losses, > and are almost impossible to eliminate in traditional programs, since > the evaluation functions are so complex. > > > > David > > > > > > It seems that alpha/beta pruning hit a roadblock a long time ago in > go. Now we have MC... which as you increase the number of samples.. > you start to get closer to an equivalent alpha/beta. But... there are > still huge groups of samples that are not checked... and if you want > to somehow prove you have the best move... how will you do it? Will > you make the sample size equivalent to the number of possible samples? > How will you do this practically? You can only state with a certain > confidence that you did make the best move and this would be bound by > our computational resources. > > > > > > > > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
