Lars wrote: > Anyone of you have similar or other experiences with the algorithm?
I use at runtime the same Bradley-Terry formulas Remí introduces in his paper. That is a huge advance compared to "naif urgency" scores because it gives a measure of how hard it was to "win" for a move candidate. But I use a much simpler offline learning algorithm: Compute the "naif" score just as standard: sc[0] = (times played)/(times played + times postponed) Use this as an a priori value of the score. Then, iterating through all the games many times, create a compensation weight CW rewarding scores winning positions with high total of concurrent scores and diminishing those won "too easily". Since the efficiency of the offline part of the program is not an issue, I make these in small steps iterating many times until they do almost nothing. Each time, I compare observed versus expected number of right guesses to see if it improves or not. sc[i] = CW*sc[i - 1] // * is elem by elem mul of a vector I guess it gets to more or less the same. Sure Remí's solution is more efficient and elegant. 2 more issues I am concerned about patterns: -------------------------------------------- MIAI URGENCY: When two (or maybe more) moves are extremely urgent, but it is not important which of the two because they are equivalent. In this case the high urgency is masked by the fact that it is divided between two moves. IMPLICITLY CHECKED URGENCY: When an urgent pattern was already on the board and was not played its score is overestimated. Imagine threatening a bamboo joint, preserving the connection may be a huge point when it saves a big group without eyespace. But it may also be worth nothing when both groups are dead. When it first appears, there is a high probability that it is big and, therefore, its urgency should be high. But, if it wasn't played, then it is probably not big. The next times it has to be considered much less urgent than the first time. Jacques. _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
