Quoting Jacques Basaldúa <[EMAIL PROTECTED]>:
I am not an MC developer, but as far as I know, UCT keeps a limited (i.e. n-ply) tree in memory and intentionally unbiasses the nodes to make the convergence faster, that does not change anything, assuming constant tree size.
You have misunderstood UCT. UCT searches to arbitrary depth, and it is actually not possible to say to what depth it does search. Also the evaluation at each node is the average winrate in the entire subtree, which make it a little hard to tell exactly what UTC does.
your play. Of course, if you increase the tree, you reach perfect play, that's not the point.
But increasing the tree every simulation *is* the point of UCT. There are perhaps three types of MC-programs. 1) 1-ply eval of all legal moves. 2) Any kind of alpha-beta/min-max like search to fixed depth (or as deep as possible using iterative deepening). 3) Best/unexplored-first search that expands the tree as necessary with no regard for depth. The difference between 2 and 3 is that 3 tend to find one winning move, whereas 2 finds the best winning move. The drawback of 3 is that it may spend a lot of time on a losing move and misses the winning move. So with infinite search 2 and 3 will find a winning move if any, but the size of the tree might be very different. Nr 1 will avoid blunders but will reach an asymptotic level of strength quickly. Currently 3 is easier to implement than 2, but if move ordering and move extensions for go become as well understood as for chess then maybe 2 is better as the authors of the Mogo team wrote in a recent mail to this list. A source of confusion is perhaps that when I and several others talks about MC-programs we think of type 3 because it so superior to type 1, but most people who tries MC for the first time do type 1 I guess, and in many academic papers about MC they seem to do something like 1. If you have not read the UCT page on sensei http://senseis.xmp.net/?UCT perhaps that can clarify things. -Magnus _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
