Hi,
Can someone explain me how to evaluate the results of CLOP? (http://www.remi-coulom.fr/CLOP/)
As a test I'm optimizing(?) the multiplicatin factors of my eval() routine using CLOP. These are 6 values. It has play 139436 games but the values under the "MAX" tab still are jumping multiple percents up- and down. Is there some value somewhere which tells me if enough games have been played?
CLOP?
Moderators: hgm, Dann Corbit, Harvey Williamson
-
jdart
- Posts: 4361
- Joined: Fri Mar 10, 2006 5:23 am
- Location: http://www.arasanchess.org
Re: CLOP?
If you are tuning evaluation parameters, and you can compute their partial derivative, the alternative is gradient descent using a surrogate measure instead of actual game results - the most popular variant of this is the "Texel method" (https://chessprogramming.wikispaces.com ... ing+Method).
If you are tuning something that takes integral values, or search parameters, or anything with a complex, discontinuous or nonlinear relationship to the eval, that is more difficult. You can either manually tune or there are some other optimization methods such as SPSA (http://www.jhuapl.edu/spsa/). I also have some experience with NOMAD (https://www.gerad.ca/nomad/Project/Home.html) and other global optimizers. Some people have used genetic algorithms but these do not usually provide fast convergence.
--Jon
If you are tuning something that takes integral values, or search parameters, or anything with a complex, discontinuous or nonlinear relationship to the eval, that is more difficult. You can either manually tune or there are some other optimization methods such as SPSA (http://www.jhuapl.edu/spsa/). I also have some experience with NOMAD (https://www.gerad.ca/nomad/Project/Home.html) and other global optimizers. Some people have used genetic algorithms but these do not usually provide fast convergence.
--Jon