Exploring and optimizing dynamic neural fields parameters using Genetic Algorithms

The Continuous Neural Field Theory introduces biologically-inspired competition mechanisms in computational models of perception and action. This paper deals with the use of Genetic Algorithms to optimize its parameters, as to guarantee the emergence of robust cognitive properties. Such properties include the tracking of initially salient stimuli despite strong noise and distracters. Interactions between the parameter values, input dynamics and accuracy of model, as well as their implications for Genetic Algorithms are discussed. The fitness function and set of scenarios used to evaluate the parameters through simulation must be carefully chosen. Experimental results reflect an ineluctable tradeoff between generality and performance.