Thank you for your ideas. The shapes Rémi is using have values as high as 143473. Of course, that would invalidate my idea. But I planned trimming the upper probability to a reasonable number, say 100 as scaling everything to a 1..100 scale (in the runtime database). I don't know yet if that makes any difference for random playouts, but I suspect it doesn't. The real application of a good distribution over the legal moves is searching, and that does not require any trimming.
Rémi, are your values the result of learning in masters games? I keep a "best quality" (= no tricks) database of urgencies and a compacted one with rounded scores and some tricks for compacting it even more which result in minor score differences. I my case, Rémi's idea is not applicable because the size of the patterns makes 9 lines ~ 50% of the board, while 40 neighbors are maximum 11% of the board (in case they all were legal moves). 50% is not really too different from rescanning everything (which is simpler). > .. but adding knowledge and high-level algorithmic improvements > is tremendously more profitable in terms of playing strength than > finding tricks to accelerate playouts. Of course. But one day you have to care about these things. If you do it well, you will never have to look at that code again. If you don't, you will have a known or ignored problem forever. The main reason for caring about doing things the right way is not having to care again. ;-) I am sure you agree because you also took the time to find the best solution for your case. Jacques. _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/