> Of course you can do put much more clever prior if you are a player and > know the subtleties of the game. >
E.g. patterns extracted from databases - but it's not enough, carefully tune the coefficients for "empty triangles" (important!) and various other importants patterns/rules (don't just keep the empirical probabilities from datasets as coefficients). I'm afraid the coefficients are very implementation-dependent unless the bandit and the random player are specified with a lot of details. You can put an exploration term, but the cases where it is needed are rare. > I did a lot of experiments on that, and even at long thinking time, no > exploration term was always better (statistically). > Well, now mogo has an exploration term - but not at all UCB-like. As said previously, I'm not the main author of this modification and I'll not explain it myself. Also, for the "main" node (the root), the bandit in mogo is different from the bandit elsewhere in the tree. More diversification is useful. But it's not a very important improvement, a few percents. Best regards, Olivier
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
