Rémi Coulom wrote: > 荒木伸夫 wrote: >> I have considered this, and I think that this may be caused by wrong >> training model. In my master thesis, I mentioned that the >> relationship between >> top 1 accuracy of move prediction and the strength of Monte-Carlo >> is not simple (I increased the number of matches to 600, and similar >> tendency appeared). Therefore, it might be wrong to use only one human >> move (top 1 move) as a positive example (such training will highten >> top 1 accuracy). We may need to use another training model... > > Unfortunately, I don't believe a usable training model exists, besides > playing plenty of games with the full MC tree search to figure out > which weights produce the best playing strength. > > A big problem is the sample distribution. Whatever patterns we use, > they are general rules with exceptions. That is to say it is always > possible to make up a weird (or not so weird) position where patterns > fail. And when a MC program is using patterns, it is naturally > attracted towards positions that are evaluated wrongly. This all (combined with the results of the study) makes me think mogo and probably the other UCT programs should be searching a little wider at long time controls. At "normal" levels they are probably very well balanced.
- Don > > Rémi > _______________________________________________ > computer-go mailing list > [email protected] > http://www.computer-go.org/mailman/listinfo/computer-go/ > _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
