On Wed, 2007-01-10 at 18:37 +0100, Łukasz Lew wrote: > On 1/10/07, Don Dailey <[EMAIL PROTECTED]> wrote: > > On Wed, 2007-01-10 at 12:12 +0100, Łukasz Lew wrote: > > > > The interesting thing is that it can do a lot more play-outs when > > > > when X is high, although it is less strong. I need to understand > > > > why. > > > > > > > > Based on the paltry data I have now it's a mistake to use X that > > > > is very high. > > > > And I would point out that the evidence is paltry - need a lot more > > games to draw any conclusions. After another day of running they > > are pretty close and when you factor in the running-time difference > > the X = 100 version may be best. Of course the evidence still is > > weak and my implementation could probably be improved. > > > > I do not understand Your experiment. > Setting X higher uses less memory while loosing some (not to much) > information. > So one shouldn't expect strength improvement. Am I right?
There are several variables that might explain the playing strength: 1. Time 2. Space 3. Algorithm Presumably my program plays perfectly because I use a perfect algorithm, but the question is how well does it play given a certain amount of time and memory (space) ? If I FIX the number of play-outs - then you are right - larger X makes the program weaker. If you fix the amount of TIME, you may get a different picture. The current results indicate that the version that uses X=100 isn't that much weaker than X=5, but X=5 takes 50 percent longer to execute. So I could run 50% more simulations with X=100 and this would almost certainly make a stronger program. The current test is indicating empirically that higher X does not hurt the search very much, only a little even given the same number of simulations. I'm going to run this test for a week if I have to - I want to get a statistically significant sample, but as it stands now, x=100 is less than 30 ELO weaker than x=5. I would test x=1 except that my machine cannot run these tests at x=1 without getting memory bound. My implementation is not memory efficient, there are many improvement I would make that would enable me to use X=1 without sacrifice. - Don > BTW > I've put the epsilon trick article on my www. > (There is also a nice description of Df-Pn algorithm inside) > > Łukasz > > > - Don > > > > > > _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
