Petr Baudis wrote: > This seems like not very productive line of argumentation unless > preceded with more exact definitions of strong.
My only claim is that "it is a hard problem". That is unobjectionable no matter how you define strong (obviously: random < strong < perfect) I can't understand why you object on this. > Actually, there is a thread about exactly this on fuego-devel In fact it is not "exactly this" it is a different approach. The post in fuego-devel tries to determine the status of each point of the board. That is not a good idea with or without MCTS because go is about trading. (Furikawari) My different approach is determining by "how many points" the simulated games are won. Only in yose the IQR becomes narrow enough to see how much territory is still in dispute. Stefan Kaitschick wrote: > If resources were no problem, the best way to estimate the score > would probably be to have an MC program find the komi that results > in a 50% winrate. Yes! That is my proposal. Saying an MCTS program is happy (60%) or unhappy (32%) does not inform too much to non-computer-go observers. If you talk about "winning rate" they understand "the probability that the program wins" which is not the case. Telling them the program is happy, but would end being happy if it had to win by 20 additional points is crystal clear. The "correct" way to determine the komi shift would be to try all possible values. Since that is expensive, my proposal estimates the komi shift from one single 20K run by studying the distribution. Of course, it is only an estimation and the other method would be more accurate. Additionally, the estimator gives a confidence interval or remembers the observer that score estimation in move 60 is a fallacy. We have to live with the two facts: * Each board position has a value. (The game satisfies the conditions of the minimax theorem.) * Pretty much every position has a value that is computationally untreatable. And this applies to human estimators as well. They only score according to established practice, which is something that is revised as new empirical evidence shows up. Scoring at move 60 is just an educated guess. (Of course people will more likely accept the guess of a 9d than the guess of a 15k.) The "cool" part is the estimator can tell you the difference when it is just making an educated guess and when most of the territory is already "sold out" with few points in dispute. Jacques. _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
