Revision history[back]
click to hide/show revision 1
Revision n. 1

Jul 01 '10 at 00:01

Rueben's gravatar image

Rueben
163

An alternative I understand works relatively well when evolving neural networks is Novelty Search ( http://eplex.cs.ucf.edu/papers/lehman_ecj10.pdf ). The short version of the algorithm is that, rather than using the traditional fitness value for determining the next generation, you chose the genotypes whose behavior is most different from the others (and from an archive of the historically most unique). Although it was designed with neural nets in mind, it should be easily adaptable to other problem spaces, even if the "behavior" of a given genome is simply some measure of distance between the other genomes that exist in the search space.

If your goal is to eventually find the global minimum, perhaps you could perform 100 generations with Novelty Search and 100 generations with traditional, fitness-based search, or any number of variations on the novelty-for-a-while/fitness-for-a-while system.

click to hide/show revision 2
converted the url to a link

Jul 01 '10 at 00:17

Rueben's gravatar image

Rueben
163

An alternative I understand works relatively well when evolving neural networks is Novelty Search ( http://eplex.cs.ucf.edu/papers/lehman_ecj10.pdf http://eplex.cs.ucf.edu/papers/lehman_ecj10.pdf ). The short version of the algorithm is that, rather than using the traditional fitness value for determining the next generation, you chose the genotypes whose behavior is most different from the others (and from an archive of the historically most unique). Although it was designed with neural nets in mind, it should be easily adaptable to other problem spaces, even if the "behavior" of a given genome is simply some measure of distance between the other genomes that exist in the search space.

If your goal is to eventually find the global minimum, perhaps you could perform 100 generations with Novelty Search and 100 generations with traditional, fitness-based search, or any number of variations on the novelty-for-a-while/fitness-for-a-while system.

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.