Re: [computer-go] RAVE in MoGo paper
Hi, 2007/10/8, Benjamin Teuber [EMAIL PROTECTED]: Hi everybody - especially Sylvain =) I'm wondering whether the formula to determine the balance between RAVE and UCT, beta = sqrt(c / 3 * parentVisits + c), has any mathematical background - or is it just a best guess for something that starts at 1 and is 1/2 after a certain number of visits? No it is just a tuning :) Another question is about the prior integration. Apparently the prior, RAVE and UCT values are three different estimators for the winning probability. So why not use the above formula for prior vs. RAVE balancing, too, instead of initializing RAVE with it? Our prior is actually classical and equivalent to a Dirichlet prior for the RAVE value. Of course we could put the prior in other ways, put I strongly believe that at this point the relevance of the prior is more important that the way you use it. Cheers, Sylvain ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] RAVE in MoGo paper
I'm wondering whether the formula to determine the balance between RAVE and UCT, beta = sqrt(c / 3 * parentVisits + c), has any mathematical background - or is it just a best guess for something that starts at 1 and is 1/2 after a certain number of visits? I guess it is simply a kind of parameter tuning. At least the constant number 3 is meaningless in the formula - we can use the following formula with c2 = c/3. beta = sqrt(c2 / (parentVisits + c2)) Another question is about the prior integration. Apparently the prior, RAVE and UCT values are three different estimators for the winning probability. So why not use the above formula for prior vs. RAVE balancing, too, instead of initializing RAVE with it? Because the prior values do not change during simulations like RAVE and UCT values. Of course there might be a more effective integration method, however we need very long time to find it. -- Yamato ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/