Thèse soutenue

Evolution de contrôleurs neuronaux plastiques : de la locomotion adaptée vers la locomotion adaptative

FR  |  
EN
Auteur / Autrice : Thierry Hoinville
Direction : Stéphane Delaplace
Type : Thèse de doctorat
Discipline(s) : Informatique
Date : Soutenance en 2007
Etablissement(s) : Versailles-St Quentin en Yvelines

Résumé

FR  |  
EN

Evolutionary robotics mainly focused on evolving neural controllers that are structurally and parametrically fixed, for the control of robots that can roll, walk, swim or fly. This approach led to the design of controllers that are well adapted to constant environments, but not adaptive to varying conditions. To tackle this issue, some researchers suggest to evolve plastic, rather than fixed, neural controllers. Our work follows this way and aims to design plastic neural controllers for legged robots subject to external perturbations, as well as possible mechanical damages. First, we propose a review of the main known forms of neuronal plasticity and their modeling. This review is mostly intended to the roboticists audience. Then, we draw a state of the art of evolving plastic neural controllers and criticize the biological realism of the developed models. On this background, we provide a first contribution centered on thedilemma of evolving both flexible and stable neural controllers. Thus,we suggest to use homeostatic constraints to stabilize CTRNNs incorporating adaptive synapses. We applied this method with success to the locomotion of a one-legged robot, confronted to an external perturbation. Finally, we present a second work based on the knowledge acquired on the biological central pattern generators (CPG) and their plasticity. In practice, we evolve neural relaxation oscillators subject to neuromodulation, for the adaptive locomotion of a modular myriapod robot, that could experience leg amputations.