I'm reading the paper by Kocsis and Szepesvari. From Figure 2 the significant imporvement of UCT over alpha-beta happens when the error tolerance is about 10%, where the improvement is a factor of 100. From the results of CGOS UCT-MC performed seemly better than this. It's propabaly due to the reason that in above paper they used randomly generated values. On the other hand MC score and the Go game values are more structured. Thus, UCT-MC performed better. But anyway the general trend of Fig.2 is probably true. The better a Go player is, the smaller the error tolerance is. In some critical situation, such as ladders, the error tolarance is almost zero. At top human player level the error tolerance is < 1%? Daniel Liu ________________________________________________________________________ AOL now offers free email to everyone. Find out more about what's free from AOL at AOL.com.
_______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
