Hello, > Perhaps I am oversimplifying, but if I understand correctly, MoGo > primarily gets its strength in 9x9 go by improving upon the random > simulations by preferring "good" moves over purely random moves during > the random game. It is not exactly that. The claim is that the improvement comes from the "sequences of moves" and not prefering "good" moves. For example, and "strange" results we get is that, if we look for the positions matching the patterns everywhere and not only near the previous move, the evaluation becomes less precise. It seems non intuitive, but it is reality :). We think that sequences of moves are far more important than the moves itself (for MC).
> Can > I conclude that those improvements to the random simulations actually > have no effect on the performance of the program? That can even have a negative impact. > But even more confusing to me is that we've tried some simple > improvements to the random program that have had no effect. The ones > that I was certain would improve performance were versions that changed > random simulations so that moves near existing stones would be preferred > over stones placed too far away from the action. Many versions have been > tried, e.g., moves that must be adjacent to some other stone, moves that > must be no more than 1 space away from existing stones, etc. Surely on > average these are going to be better moves than purely random moves -- > or is this, indeed, the flaw in my logic? Shouldn't these versions > outperform the purely random versions? We tried exactly the same things at the very beginning of MoGo. The improvement on self play was significant, but not against gnugo. However, this was the beginning of understanding the idea about sequences I mentionned above. So historically this kind of experiments brough improvements in MoGo latter. Sylvain _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
