>> In my old Goliath program I used positions as hashcodes to lookup a >> move/value combination because it was much easier to program than >> trees :)
>I think this is the better solution and it's probably easier too. You probably need both an automatically generated system (self-learning) and a manual override. You need the automatically generated system because computers can fill in much larger sections of the tree than any human possibly can. The quality can easily be made higher than your engine can create over the board. In most domains there are problems that defy computer solution. Even Chinook struggled with one particular checkers opening and required manual instruction despite having deep search, carefully tuned heuristics, and an exhaustive library of 9-piece endgames. IIRC, it was only with the 10-piece endgames that Chinook was able to see through the variations from the root. For Go, Fuego uses hand-coded openings and Mogo uses automated learning. Martin's recent match report makes clear that both systems have relative strengths. Here is a tidy illustration of both strength and weakness of manual book preparation: Martin had to change an opening that had long been in Fuego's book because Fuego had done badly with it. For the other side, Mogo was able to play an amazing 13 book moves at the start of one game, but then found that the position was not as good as it thought. Pebbles is doing both. Maybe someday it will do both well. :-) Brian _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
