Parameter tuning with multi objective optimization

Discussion of chess software programming and technical issues.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
Posts: 686
Joined: Thu Mar 03, 2011 3:57 pm
Location: Germany

Re: Parameter tuning with multi objective optimization

Post by tpetzke » Wed May 10, 2017 12:27 pm


I use genetic algorithms for some years now to tune iCE and for me it works in general. In general means it improves a not tuned evaluation quiet a lot. The better the evaluation gets the more likely further runs will not be able to improve it further or might even come up with slightly worse weights.

Fitness functions that don't factor in game play have not worked for me. I always needed to play matches. However I use a set of test positions to filter out population members that have an absurd fitness (e.g. if the value of the queen is less than a knight it will not score very well in the set) just to save compute time. I also use it to decide draws. As engines that score better in the set have a higher chance of survival the final genom is also solving the set better but is not overfitted towards it (e.g. sacrificing playing strength to solve 5 additional positions).

The computational effort for a single run (1000 generations, population size 128 - 256) in my framework is very high (2 to 3 weeks) so I tried a lot of different cheaper fitness functions unfortunately without any success.

======= - iCE Chess Engine

Post Reply