The success of the UCT-MC method, especially in the 19x19 Go game, surprises many people. It has been?mostly atributed to?the special features of the UCT algorithm. These features does play an essential roles They made it possible for the evaluation method of Monte-Carlo to work. However, the success of the?UCT-MC is largely due to the nature of the Monte-Carlo(MC) evaluation method. I'll explain this.
First, let's give a more precise definition of the MC evaluation. That is MC simulation is a random function. It's variable is the random playing sequence. The distribution of the MC random function approximate the structure of the evaluation space of the Go game. Now we need the definition of the evaluation space. The evaluation space?consists of all possibe final score and move sequence pairs. The success of the MC evaluation can be explained by the following example. Let's assume we have 4 magnetic balls placed on a surface. Two painted white and two painted balck.?If we throw a 5th magnetic ball at them, what's the probability of the 5th ball stick with a white ball? An alpha-beta search will answer the question by trying out all possibilities. A MC evaluation will answer the question by doing simulations.?As the 4 magnetic?balls?are concerned there isn't much difference in computation time between the two methods. However, when the number of the colored magnetic balls becomes very large, the advantage of the MC method becomes?better factorially. In another word the MC evaluation turns the discrete problem into an continuoues area problem. DL? ? ________________________________________________________________________ More new features than ever. Check out the new AOL Mail ! - http://webmail.aol.com
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
