Joel Veness wrote: > Hi Don, > > That is an interesting idea, and I don't see anything wrong with it in > principle. > > However, given how slow Goanna is these days, I think I would probably > gain more by spending 1/10 of the time needed for that idea on simple > optimization! :) >
Yes, I don't use the idea myself, I just do the simple test for 0.95. However sometimes doing the "more correct" thing actually is just as simple. I think this would be just a little piece of code that would actually replace the conditional branching and perhaps even make the code look cleaner. If I ever get around to implementing it myself, I'll let you know. - Don > Joel > > On Tue, Feb 5, 2008 at 2:44 AM, Don Dailey <[EMAIL PROTECTED]> wrote: > >> Joel Veness wrote: >> > Hi Nick, >> > >> > Goanna (agog) timed out annoyingly in that game against GNU. >> > >> > I have since implemented a rule: "if after some number of samples you >> > have a winning probability that is very close to 1.0, just make the >> > best move right away". There is no need to spend so long thinking in >> > these ridiculous endgame positions. I made some other changes so that >> > passing is more highly favored in certain situations, so hopefully >> > this will be the last time Goanna forfeits a game due to time. >> > >> My rule is that if the score if over 0.95 or under -0.95 I allocate >> some fraction of the time I normally would to that move. I think my >> fraction is 1/10. I always play the move with the most samples, but >> for this I make sure than the highest scoring move (even with low >> samples) fit's this window. >> >> I think technically these kinds of algorithms should be applied >> gradually, so some function could be applied to the score to determine >> how soon to stop. The function should not be linear but should be >> such that very little reduction is applied unless the score is extremely >> high or low. For instance is the score is 80% you might benefit >> from a slight reduction, but not much. >> >> To find the right numbers, you would probably need to analyze a LOT of >> data and see how often you lose (and why) when the score is fairly >> high. Probably not worth the effort! >> >> This function should be fitted to your general time control algorithm >> too. If you have rules to shorten the search, you should definitely >> balance this by being more aggressive about time allocation earlier in >> the game. It would probably be very difficult to find a good >> balance but the right algorithms should cause you to lose some games and >> win some games but to win more than you lose. If your are too >> conservative it might seem ok, but that causes you to lose games too, >> you just don't notice it as much. >> >> - Don >> >> >> >> >> >> >> >> > I guess that is what I get from only testing on CGOS. The small time >> > increment CGOS gives makes a big difference with the high latency I >> > get due to being based in Sydney, and to the best of my knowledge, KGS >> > doesn't do the same. >> > >> > Joel >> > >> > On Tue, Feb 5, 2008 at 1:32 AM, Nick Wedd <[EMAIL PROTECTED]> wrote: >> > >> >> Congratulations to the three winners of yesterday's KGS bot tournament >> >> (there was a tie in the Formal division, so three winners). My report >> >> is at http://www.weddslist.com/kgs/past/35/index.html >> >> >> >> Nick >> >> -- >> >> Nick Wedd [EMAIL PROTECTED] >> >> _______________________________________________ >> >> 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/ >> >> > _______________________________________________ > 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/
