>I haven't tried it, but (with the computer chess hat on) these kind of >proposals behave pretty badly when you get into situations where your >evaluation is off and there are horizon effects.
In computer Go, this issue focuses on cases where the initial move ordering is bad. It isn't so much evaluation/horizon. For example, at some node moves A, B, C, and D are seen as the best move, but they all fail. So perhaps you invest 2000 moves in each, with a 40% win rate overall. Now you come to move E, which is best. In fact, E is a refutation, with a 70% win rate. The question is: how long will it take before this node is considered to be refuted? The problem is that we started the node badly. When E is first explored, the node has 3200 wins, 4800 losses. Even with a 70% win rate, E needs 4000 trials (= 2800 wins, 1200 losses) just to get back to a 50% win rate. But if you had explored E first, then this node would have been considered refuted after maybe a few hundred trials. There is literally a factor of 100 difference in search efficiency. This situation arises seems odd, but arises frequently. For example, if E is a winning move now, but becomes a losing move if played in the future. (For example, E is in a capturing race that is currently winning for the side to move first.) Unless E is tried early in the node's lifecycle, the RAVE score for E is very low. >> I recall a paper published on this basis. A paper presumably about >> CrazyStone: Efficient Selectivity and Backup Operators in Monte-Carlo >> Tree Search. >I'm reasonably sure this did not include forgetting/discounting, only shifting >between average and maximum by focusing simulations near the maximum. Yes, but emphasizing the maximum is equivalent to forgetting. E.g., in my example, giving E a larger weight is equivalent to giving A, B, C, and D a lower weight (== forgetting). _______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
