Thèse soutenue

Relations entre plasticité synaptique et régularité des codages en neuro-évolution

FR
Auteur / Autrice : Paul Tonelli
Direction : Stéphane DoncieuxJean-Baptiste Mouret
Type : Thèse de doctorat
Discipline(s) : Informatique
Date : Soutenance en 2012
Etablissement(s) : Paris 6

Mots clés

FR

Mots clés contrôlés

Résumé

FR  |  
EN

The evolution of artificial neural networks or neuro-evolution is able to generate networks capable of solving non-trivial tasks such as makin polyarticulated robots walk or enabling mobile robots to navigate autonomously. However, in most cases, the evolved controller may only solve the problemswhich have already been encountered during the evolutionary process, and the controller cannot adapt online to cope with new situations. Our work establishes the link between two domains of neuro-evolution previously studied independently. On the one hand, synaptic plasticity mechanisms, used as building block in these algorithms to create networks that adapt online and on the other hand, the generative encodings promoting regularity in the createdneural networks. The first contribution of our work is to show that the combination of these two tools provides the ability to generate networks that can adapt online to unknown situations, while the majority of other methods can not: the networks are often only able to "switch" between cases presented during evolution (a phenomenon that can be compared to overlearning) or can only be robust to changes in the environment without any qualitative change in their behavior. Generative encodings promoting regularity, on the contrary, where the same structures are applied to a set of neurons, make it more difficult for the algorithm to overspecialize the evolved networks. Our second contribution solves another problem encountered when evolving neuralnetworks using generative encodings promoting regularity: these encodings generate networks where the behavior of "bundled" neurons is similar, which does not allow them to single out the behavior of any neuron. The addition of synaptic plasticity mechanisms relaxes this constraint by allowing the network to learnin a set of behaviors from a common rule, even if these behaviors are different from one another.