Memory is an issue with these programs, since they build tree's and > > maintain them in memory. So none of these programs can think for > more > than a few minutes per move. > > > dimwit gets around this problem by increasing the number of visits > required before a node is added to the UCT structure as the structure > gets big. This works fairly well, and I can let it run for hours on a > position. However, the standard boost of the uncertainty in the score > by sqrt(log(total_visits)) grows way too slow, and the main line goes > way too deep, without giving other moves a fair chance. Are you saying this is an imbalance caused by increasing the number of visits required or a problem in general? If you were given infinite memory would the behavior be cured?
> > I suspect that for very long time controls we would be better off > turning UCT (with, say 10K playouts) into an evaluation function and > then using alpha-beta on top of it. > > Álvaro. This is very interesting to me. Not the memory management part, but the fact that you believe the tree is not being grown optimally (if that is what you are saying.) I think that is probably true for all of our programs. My theory is that programs always tend to be optimized for the current hardware because we run empirical tests on them and keep what seems to work the best. I can imagine that in 5 years, we might "thicken" the tree a bit simply because it seems to work better. But we might find that if we dust off the old computers we had 5 years ago this new improvement would not work so well. This probably just means we are not doing it quite right in the first place. We are tuning our programs for the "sweet spot" which happens to be whatever computer we are running on at the time. - Don > ------------------------------------------------------------------------ > > _______________________________________________ > 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/
