>2) Many go program authors have stated that "play to maximize wins" >is stronger than "play to maximize points". I think this is because >their evaluation functions are imperfectly optimistic--the program >counts points that future play does not deliver.
You could be right, because really this is something that no one is sure of. But I will give a 90% confident reply that you are wrong, and using point differential in MCTS *should* be weaker than using winning percentage. The problem with point differential is the signal to noise ratio. The result of a single trial could have a standard deviation of 100 points, so if you are trying to identify the move with the highest point differential then you have a lot of statistical noise to work though. Winning percentage has only two outcomes, so the standard deviation is bounded. And it seems (experimentally) that the difference between best and second-best plays based on winning percentage (the "signal") is a larger fraction of the standard deviation. Now, it could be that this is just an implementation artifact, and if someone wrote a better estimator of point differential then we would see different things. And that's possible. But then every MCTS Go program does use winning percentage, so we believe that this is implied by the domain. Brian _______________________________________________ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
