Evolving Variants of Neuro-Control Using Constraint Masks

The search for variants of effective neural behavior is a major requirement for the identification of novel neuro-dynamical control principles. Evolutionary algorithms are successfully used to search for such controllers. But neuro-evolution tends to find similar, well performing solutions when run multiple times, instead of many, perhaps also weaker performing, but neuro-dynamically highly interesting variants. Furthermore, variants only develop by chance, so that a systematic exploration of different neural control strategies is difficult. With the ICONE method the search space can be shaped by so-called constraint masks (CM) to bias the evolving networks towards specific configurations. On the basis of an animat experiment we demonstrate that the number of evolved distinct variants can be significantly increased using different CMs.