On Thu, Apr 05, 2007 at 10:49:05AM +0100, Jacques Basaldúa wrote: > > Many users feel stolen by UCT programs. I have read that > in the KGS chatrooms. Normal users do not count with > +/- 0.5 point precision. They have the impression the > program blundered and they caught up. But when the > program counts, surprise!, it wins by 0.5 points Chinese. > The users were thinking Japanese even if they accepted > Chinese rules. In fact, they did not have the choice. > They get the impression the game was stolen by > "technicalities" after they saw the engine blunder many > times.
As I have posted here before, this is easy to avoid, by counting the result of teh game not as +/- 1, but as +/- (1+epsilon), where epsilon is so small that it can not change the order of the averaged-out results. If you first normalize the matrgin of victory to 0..1 by dividing by boardsize, and then divide by by the number of playouts you do, you get a suitably small epsilon, that only affects the sorting of games that get as many wins. Of those the program will prefer the moves that win by a larger margin. This should not have any effect on the strength, because even one playout difference is greater than all teh epsilons summed up, but in the endgame, when all moves lead to a certain victory, the program prefers a win by a larger margin. My Halgo does that (in a otherwise regular MC), and the feel of the endgame is clearly more reasonable than without. -H -- Heikki Levanto "In Murphy We Turst" heikki (at) lsd (dot) dk _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
