Wednesday, February 16, 2005


Supergenes and competence

Javarunner continues to post about JGAP and evolutionary computation here, but digging below the surface shows an interesting post on supergenes. Supergenes in JGAP aggregate variables that are interrelated either from inspection of the set of equations being optimized or intuition. The supergene page claims the following:
The supergene conception can significantly (up to 3 times in the most boundary cases) increase the evolution speed. It can also increase the accuracy of solution for the small populations. However these effects are dependent from the chosen population size and maximal number of iterations. While in some cases the use of supergenes can be definitely sensible, we would recommend to try both supergene and non-supergene versions for your specific task.
Interestingly, IlliGAL work has demonstrated significant speed ups (sometimes going from exponential scalability to subquadratic) in hard problems in similar way, except IlliGAL competent GAs require no prior knowledge of the problem being solved or the variable interrelationships. This approach works, because initial results are used to determine which variables are interdependent, and then "supergenes" are automatically constructed to solve the problem quickly.

Similar automatic linkage detection or supergene detection could be built into JGAP and other traditional GAs so they automatically give faster, better solutions to hard problems without as much prior problem knowledge or understanding.

can you put up a link or reference to this work at illigal? it sounds quite handy.
The best overview of GA competence is in the book The Design of Innovation available in libraries and online booksellers. Pelikan's work on hBOA, Yu's work on DSMGA, Harik's work on ECGA are all available as individual papers from the tech reports section of Some code is also online for those who want to give it a shot.
Post a Comment

<< Home

This page is powered by Blogger. Isn't yours?