On Thu, 2008-10-16 at 08:47 -0400, Michael Williams wrote: > What if you use a faster, lower quality RNG? How much do the numbers change?
I don't even know if my current generator is high quality - I'm using the standard RNG library that comes in the IBM java distribution. Are these standardized between java implementation? I have another package of RNG's that I could try that are considered high quality and include mersene twister. I could also implement some trivial bad RNG just to see what happens. I used a very simple low quality RNG in one of early programs (because I needed to incorporate it into a small handheld device.) Subjectively, I could not see that the quality of play was affected but of course this would have to proved. The main side-effect was that you would get repeated games if you self-played enough of them because it had a low period. (This is true of ANY PRNG but you wouldn't easily be able to demonstrate it.) I believe such a generator would likely fail a test like this, for instance it might NEVER play D4. I think I might give that a try later to see what happens. - Don > > Don Dailey wrote: > > Update: > > > > 4173 runs of 100,000 playouts from opening at 0.5 komi > > > > mv: D4 count: 3 percent: 0.0719 > > mv: D5 count: 447 percent: 10.7117 > > mv: E5 count: 3723 percent: 89.2164 > > > > > > 0.959 percent fall outside the following range ... > > > > score: lo, med, hi -> 0.52031 0.52433 0.52835 > > nodes: lo, med, hi -> 11092602.0 11105554.0 11119436.0 > > > > > > > > > > On Thu, 2008-10-16 at 00:14 -0400, Don Dailey wrote: > >> I have some interesting statistics on the simple go program at 0.5 komi > >> from the starting position. > >> > >> I'm running numerous 100,000 game samples and tracking the statistics to > >> see what kinds of variation I get in scores and nodes. > >> > >> After 1748 runs I see that less than 1 percent of the games score lower > >> than 0.52027 or higher than 0.52823 when doing 100,000 game playouts. > >> > >> So if you get scores outside this range you probably do not have a > >> conforming program as this is expected to happen less than 1% of the > >> time. > >> > >> But what I really found interested is that only 3 moves (when accounting > >> for transformations) were chosen from the opening position. E5 was > >> chosen 85% of the time, and most of the remaining time D5 or > >> equivalent. Only 1 time was some other move chosen other than these > >> two and it was D4. > >> > >> I wonder how long before it would chose A1? Probably a very long time > >> indeed! > >> > >> So if your bot chooses a move other than E5 or D5, there is a very good > >> chance it is not conforming to our specification of a generic MC player. > >> > >> --- > >> > >> > >> 1748 runs > >> > >> mv: D4 count: 1 percent: 0.0572 > >> mv: D5 count: 200 percent: 11.4416 > >> mv: E5 count: 1547 percent: 88.5011 > >> > >> > >> 0.915 percent fall outside the following range ... > >> > >> score: lo, med, hi -> 0.52027 0.52434 0.52823 > >> nodes: lo, med, hi -> 11093084.0 11105628.5 11119815.0 > >> > >> > >> _______________________________________________ > >> 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/ > > _______________________________________________ > computer-go mailing list > [email protected] > http://www.computer-go.org/mailman/listinfo/computer-go/
signature.asc
Description: This is a digitally signed message part
_______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
