>it seems a little strange to me that all >outcomes below the current position are just added up.
Well, the weight is a second-order effect. The ratio of wins to losses is primary. Any move that wins at any time in the future does suggest that move might be winning now. Changing the balance between early and late has pros and cons. If later moves have lower correlation to game-theoretic win/loss then a lower weight helps to reduce their effect. But if the correlation is positive then you do want to include such observations. Lower weights have the side-effect of increasing the RAVE UCB value; that is, if the weight is less than 1.0 then the UCB will be larger at the same number of RAVE trials. This will cause actual UCT trials to be allocated to moves. It is not clear if that is beneficial. Fuego uses a lower weight for distant moves than for nearby moves. I suspect that isn't much better than using uniform weight. I am hope that Martin or Markus will comment.
_______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
