TalkChess.com
Hosted by Your Move Chess & Games

Author Message
Don Dailey

Joined: 29 Apr 2008
Posts: 5106

Post subject: Re: CLOP for Noisy Black-Box Parameter Optimization    Posted: Wed Sep 07, 2011 6:09 pm

Rémi Coulom wrote:
Don wrote:
 Rémi Coulom wrote: Comments and questions are welcome. Rémi

Larry and I are playing around with CLOP using clop-gui.

In the Max tab, there appears the columns Mean, Max-1, Max-2 and apparently as many columns as there are numbers of parameters to be tuned.

However nothing ever appears in those columns after a day of running. What do those columns mean (no pun.)

I explained earlier in this thread: it is a feature I did not implement. Only max-1 has a meaning: it is the maximum of the quadratic regression, if it exists. It sometimes does not exist. Mean is the weighted mean of all samples. I recommend using parameter values in the "Mean" column for your parameters. They usually perform better according to my artificial experiments.

 Quote: In the paper you mention that this is supposed to converge - so given enough games and time will this eventually stop running? I told Larry that even with a very efficient algorithm it is likely to take millions of games to see values that improve on already well selected values.

It does not converge in a finite amount of games. It converges in the sense that expected simple regret decreases in time, and can be made arbitrarily small with a big enough number of games. Note that I do not have a mathematical proof of this convergence, but I am convinced it does converge if the function to be optimized is continuous.

In the paper I ran experiments on artificial problems. It gives an idea of how close to optimal parameters you get with a given number of games. Around 50% winning rate, one Elo point is about 0.0014. For the Log problem in one dimension, it takes about 10k games to have 0.001 expected simple regret (so, less than one Elo point). For the 5-dimensional problem, it takes a few hundred thousand games to reach 1-Elo expected simple regret (that's 0.2 Elo per parameter).

The magic of CLOP is that it approaches optimal performance faster than it can be measured. Simple regret is smaller than the confidence interval around the estimated win rate.

Rémi

Thanks for explaining that Rémi. I have a script for chess if anyone wants to use it for chess. The script does not have a legal move generator for simplicity and therefore draw and win detection is a bit of a kludge, but it works well.
 Display posts from previous: All Posts1 Day7 Days2 Weeks1 Month3 Months6 Months1 Year Oldest FirstNewest First
Subject Author Date/Time
Rémi Coulom Thu Sep 01, 2011 10:32 am
Joona Kiiski Fri Sep 02, 2011 8:52 am
Rémi Coulom Fri Sep 02, 2011 9:45 am
Marco Costalba Fri Sep 02, 2011 5:34 pm
Zach Wegner Fri Sep 02, 2011 4:44 pm
Rémi Coulom Fri Sep 02, 2011 7:58 pm
Rein Halbersma Fri Sep 02, 2011 10:07 pm
Rémi Coulom Sat Sep 03, 2011 9:24 am
Rein Halbersma Sat Sep 03, 2011 5:01 pm
Rémi Coulom Sat Sep 03, 2011 5:28 pm
Andres Valverde Sat Sep 03, 2011 10:28 am
Rémi Coulom Sat Sep 03, 2011 10:40 am
Peter Österlund Sat Sep 03, 2011 6:54 pm
Rémi Coulom Sat Sep 03, 2011 7:34 pm
Ernest Bonnem Sun Sep 04, 2011 4:49 pm
Don Dailey Wed Sep 07, 2011 4:57 pm
Rémi Coulom Wed Sep 07, 2011 5:47 pm
Re: CLOP for Noisy Black-Box Parameter Optimization Don Dailey Wed Sep 07, 2011 6:09 pm
Rémi Coulom Wed Sep 07, 2011 6:26 pm
F. Bluemers Wed Sep 07, 2011 6:39 pm
Rémi Coulom Wed Sep 07, 2011 8:19 pm
Rémi Coulom Thu Sep 08, 2011 4:44 pm
Larry Kaufman Fri Sep 09, 2011 3:36 am
Rémi Coulom Fri Sep 09, 2011 6:11 am
Larry Kaufman Fri Sep 09, 2011 2:43 pm
Rémi Coulom Fri Sep 09, 2011 7:47 pm
Marco Costalba Sun Nov 06, 2011 8:24 am
Rémi Coulom Sun Nov 06, 2011 8:30 am
Marco Costalba Sun Nov 06, 2011 8:50 am
Peter Österlund Sun Nov 06, 2011 1:11 pm
Marco Costalba Sun Nov 06, 2011 4:53 pm
Rémi Coulom Sun Nov 06, 2011 5:14 pm
Peter Österlund Sun Nov 06, 2011 11:11 pm
Rémi Coulom Mon Nov 07, 2011 11:18 am
Rémi Coulom Sun Nov 06, 2011 5:05 pm
Rémi Coulom Fri Sep 09, 2011 7:52 pm
Rémi Coulom Sat Sep 10, 2011 9:49 am
Michel Van den Bergh Sun Sep 25, 2011 8:18 am
Rémi Coulom Sun Sep 25, 2011 8:47 am
Michel Van den Bergh Mon Sep 26, 2011 10:55 am
Rémi Coulom Mon Sep 26, 2011 1:40 pm
Michel Van den Bergh Tue Sep 27, 2011 7:54 am
Rémi Coulom Tue Sep 27, 2011 8:39 am
Michel Van den Bergh Thu Sep 29, 2011 2:04 pm
Rémi Coulom Thu Sep 29, 2011 3:38 pm
Daniel Shawul Mon Oct 17, 2011 2:03 am
Michel Van den Bergh Mon Oct 17, 2011 7:51 am
Rémi Coulom Mon Oct 17, 2011 7:56 am
Rémi Coulom Mon Oct 17, 2011 8:02 am
Rémi Coulom Mon Oct 17, 2011 8:11 am
Daniel Shawul Mon Oct 17, 2011 9:54 am
Rémi Coulom Mon Oct 17, 2011 11:10 am
david nash Wed Oct 05, 2011 12:19 am
Edward Yu Fri Oct 07, 2011 11:58 am
Rémi Coulom Fri Oct 07, 2011 1:17 pm
Edward Yu Fri Oct 07, 2011 4:01 pm
Rémi Coulom Fri Oct 07, 2011 4:30 pm
Edward Yu Sat Oct 08, 2011 3:59 am
Rémi Coulom Sat Oct 08, 2011 7:49 am
Edward Yu Tue Oct 11, 2011 4:04 pm
Rémi Coulom Tue Oct 11, 2011 4:22 pm
Daniel Shawul Tue Oct 11, 2011 4:49 pm
Rémi Coulom Tue Oct 11, 2011 5:27 pm
Michel Van den Bergh Sat Oct 15, 2011 2:51 pm
Rémi Coulom Sun Oct 16, 2011 6:30 am
Michel Van den Bergh Sun Oct 16, 2011 8:10 am
Rémi Coulom Sun Oct 16, 2011 9:02 am
Michel Van den Bergh Sun Oct 16, 2011 3:10 pm
Rémi Coulom Sun Oct 16, 2011 7:25 pm
Michel Van den Bergh Sat Oct 08, 2011 6:00 am
Rémi Coulom Sat Oct 08, 2011 7:53 am
Rémi Coulom Sat Oct 08, 2011 8:27 am
Ferdinand Mosca Tue Oct 11, 2011 8:22 am
Rémi Coulom Tue Oct 11, 2011 1:38 pm
Ferdinand Mosca Sun Oct 23, 2011 7:39 pm
Rémi Coulom Sun Oct 23, 2011 8:55 pm
Ferdinand Mosca Sun Oct 23, 2011 10:47 pm
Jonatan Pettersson Fri Nov 04, 2011 9:42 am
Rémi Coulom Fri Nov 04, 2011 9:58 am
Jonatan Pettersson Fri Nov 04, 2011 10:17 am
Rémi Coulom Fri Nov 04, 2011 10:30 am
Jonatan Pettersson Fri Nov 04, 2011 10:44 am
Rémi Coulom Fri Nov 04, 2011 10:56 am
Jonatan Pettersson Fri Nov 04, 2011 11:04 am
Rémi Coulom Fri Nov 04, 2011 11:56 am
Jonatan Pettersson Fri Nov 04, 2011 4:16 pm
Jonatan Pettersson Fri Nov 04, 2011 5:48 pm
Sergei S. Markoff Fri Nov 04, 2011 3:02 pm
Rémi Coulom Sat Nov 05, 2011 7:45 pm
Jonatan Pettersson Sat Nov 05, 2011 10:49 pm
Ilari Pihlajisto Sat Nov 05, 2011 11:17 pm
Jonatan Pettersson Sun Nov 06, 2011 12:06 am
Rémi Coulom Sun Nov 06, 2011 8:00 am
Jonatan Pettersson Sun Nov 06, 2011 10:20 am
Rémi Coulom Sun Nov 06, 2011 5:10 pm

 Jump to: Select a forum Computer Chess Club Forums----------------Computer Chess Club: General TopicsComputer Chess Club: Tournaments and MatchesComputer Chess Club: Programming and Technical DiscussionsComputer Chess Club: Engine Origins Other Forums----------------Chess Thinkers ForumChess Players ForumForum Help and Suggestions
You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum