Abstract EN:

Evolutionary algoritms have been successfully used to generate controllers for many robots. However, they struggle to design complex artifacts when the fitness is unable to explicitly guide the process. In this thesis, we draw the hypothesis that these problems originate from the use of a single selection pressure, whereas living organisms are subject to many ones. We investigate here th euse of multiobjective evolutionary algorithms to create such multiple gradients in order to evolve neuro-controllers. We first describe how hypotheses about potential intermediate steps can be used by defining a multiobejctive optimization problem in which each objective corresponds to a sub-task. In the lack of any selection pressure, it is also possible to add an objective which encourages an efficient exploration of the neighborhood of current candidate solutions. We consider several possibilities to instantiate this concept for the evolution of neural networks and we conclude that maintaining the diversity of the behaviors, instead of the diversity of the genotype or the ohenotype, is an efficient way to override the deceptiveness of a fitness function. Last, we show that exaptations can be favored by applying a selection pressure on some modules of the generated neural-networks, possibly linked to genotypic modules. We tested these methods on the evolution of neural networks to compute a Boolean function and to control a light-seeking robot. They seem to be applicable to a wide range of evolutionary robotics problems, from complex locomotion to behavior control.