Hi Hiroshi, Why do you call it a static bonus?
If n increases with the number of simulations, the effect of the bonus term still fades away. Perhaps the interesting part is in fading out more slowly than with ordinary priors (i.e., by 1/sqrt(n) instead of 1/n)? BR, Erik BTW Nicks original message never arrived in my gmail account :-( On Sat, Nov 22, 2014 at 6:53 PM, Hiroshi Yamashita <y...@bd.mbn.or.jp> wrote: > Delayed congratulations to AyaMC, winner of last Sunday's KGS bot >> > > Thank you for the tournament and report, Nick! > > Recently I got +100 Elo from selfplay by adding static bonus in UCB. > I used Ikeda's paper method. > > Effciency of Static Knowledge Bias in Monte-Carlo Tree Search > Kokolo Ikeda and Simon Viennot, CG2013 > > UCB is like this. > > UCB = w/n + C * sqrt( log(N) / n ); > RAVE = Rw/Rn + C * sqrt( log(N*175) / (N*0.48) ); > beta = Rn / (Rn + n * (W1 + W2 * Rn)); > > UCB_RAVE = beta*RAVE + (1-beta)*UCB + G*log(1+gamma)*sqrt( K / (K + n)); > > n : child nodes > w : child wins > Rn: child nodes (Rave) > Rw: child wins (Rave) > N : sum of children's nodes > gamma : child gamma from MM > C = 0.31 > W1 = (1.0 / 0.9) > W2 = (1.0 / 20000) > K = 600 > G = 0.01 > > Regards, > Hiroshi Yamashita > > _______________________________________________ > Computer-go mailing list > Computer-go@dvandva.org > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
