On Dec 18, 2007 4:21 PM, Don Dailey <[EMAIL PROTECTED]> wrote: > > > Chris Fant wrote: > >>> 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 thought his point was that with an alpha-beta layer on top of the > > UCT layer, you can do much longer searches because you are throwing > > away the large UCT tree after each evaluation of an AB tree node. > > > I'm not sure he proposes this as a solution to the memory problem or > whether he believes it creates a tree with a better shape.
Actually, it's the latter. The big problem with UCT in my opinion is that it uses the exploration rule for two things: 1) visiting more probable moves more frequently, and 2) mixing the results of playouts to form a score that can be propagated up the tree. The UCB rule satisfies both purposes well enough when you are searching a few tens of thousands of simulations, but as you go to longer time controls (or, equivalently, faster hardware) you'll find that you still want to spend some time analyzing that queen sacrifice (sorry for the chess analogy), but if it results in disaster, that shouldn't pollute the score that I propagate up the tree! In UCT both notions are tied together, but as you gain more confidence in your search, you should converge to the score-backup rule of just returning the score of the best move. Once you get to that point, you can start using alpha-beta for pruning. Now it's time to get to work on dimwit to prove that there is some truth behind these reasonably sounding ideas. As John once said, our jobs are getting in the way of our go programming. ;) Álvaro.
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
