I recall looking at this algorithm (ROCK*) some time back. It was originally developed for robotics applications.
Looks like Lyudmil Antonov and Joona Kiiski have adapted it for chess:
https://github.com/lantonov/Rockstar
I don't recall this being discussed here. It kind of looks like a CLOP replacement at first glance.
--Jon
ROCK* black-box optimizer for chess
Moderators: hgm, Rebel, chrisw
-
jdart
- Posts: 4366
- Joined: Fri Mar 10, 2006 5:23 am
- Location: http://www.arasanchess.org
-
lantonov
- Posts: 216
- Joined: Sun Apr 13, 2014 5:19 pm
Re: ROCK* black-box optimizer for chess
I made this adaptation in Perl a couple of years ago. There was a slight discussion (mostly me appealing for help in development, code revision, etc) in Fishcooking forum https://groups.google.com/d/msg/fishcoo ... EJyLPJGwkJ.
My tests showed that it was able to converge to some pre-defined values which, however, does not mean automatically that it will gain Elo. I have not played games between master and tuned to show Elo gain / loss.
One of the problems in testing for convergence was the instability of the covariance matrix: with many iterations it kept getting either infinitely small or infinitely large. I guess that this is due to the non-convexity of the objective function which leads to the appearance of non-positive definite matrices. If this is the case, than instability can be possibly mitigated in a manner similar to the transformation of the Hessian matrix in adaptive SPSA.
My tests showed that it was able to converge to some pre-defined values which, however, does not mean automatically that it will gain Elo. I have not played games between master and tuned to show Elo gain / loss.
One of the problems in testing for convergence was the instability of the covariance matrix: with many iterations it kept getting either infinitely small or infinitely large. I guess that this is due to the non-convexity of the objective function which leads to the appearance of non-positive definite matrices. If this is the case, than instability can be possibly mitigated in a manner similar to the transformation of the Hessian matrix in adaptive SPSA.