On Nov 25, 2012, at 5:31 PM, Hideki Kato <hideki_ka...@ybb.ne.jp> wrote:
> Zen had chances to draw in some games in Black but selected much > risky (actually losing) moves. I guess this caused by implementing > draws by adding a third value, 0.5, to UCB. To play a draw > move, all better looking moves (by prior) have to be refused, or proved > worse than 0.5. This could take pretty long time in some positions. > #Faking komi to 6.5 might help but a better solution possible? I think there is naturally two objectives the bot can go for: • Maximize chances of winning • Maximize chances of not losing I can think of a few candidate ways to handle the dual goals. To simplify the bullets below, I'll use an input win rate of W. The most logical choice would be the win rate of the last move played in the game, but there may be less noisy alternatives. 1. Use both metrics in separate threads 2. if W > x, metric = P(win) else metric = P(not lose) 3. Weighted sum. metric = f(W)*P(win) + (1-f(W))*P(not lose). I would compare f(W)=W and f(W)=0.5, but more complex options may work better than both of those. 4. If all moves have P(not lose) < x, metric = P(not lose). Otherwise, restrict moves to P(not lose) >= x, then select based on metric=P(win) Items 1-3 can be applied to tree search. Items 2-4 can be used for final move selection. _______________________________________________ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
