Open-ended evolution of neural models exploiting behavioral consistency

Charles Ollion

ISIR, Université Pierre & Marie Curie / CNRS, France

Evolutionary algorithms have been used to tune parameters of neural models, or even to find neural structures. However it can be challenging to define a fitness function that evaluates how well the model is performing. The fitness functions that are currently used reward models that behave as closely as possible as the one the researcher wants. I will introduce a new method to design fitness functions that do not reward models that have a specific behavior, but rather models that have a coherent behavior in different contexts. For instance, they are rewarded if resistant to noise, or if they exhibit certain properties. This evolutionary method is more open ended and provides with a diversity of models with different behavior - instead of a specific behavior. The method is applied to action selection and visual attention models.

EvoNeuro: AbstractOllion (last edited 2012-07-06 08:57:41 by BenoitGirard)