EVOLUTIONARY IDENTIFICATION OF ACTIVE PARTICLE SYSTEMS

Bogdan Stanciulescu, Jean Louchet

Ecole Nationale Supérieure de Techniques Avancées

32 boulevard Victor

75739 Paris cedex15, France

e-mail: stanciul@ensta.fr

http://www.ensta.fr/~stanciul

ABSTRACT

This paper presents how it is possible to introduce active motricity into particle-bond systems used in applications such as image animation. We chose to add into some neural network capabilities over the classical approach, in order to obtain a system able to model a larger class of behaviour. Therefore a new type of binary bond enriched with a neural-based command ability is proposed and tested in this paper. This “active” bond acts like a controlled muscle in order to produce motricity.

An Evolutionary Strategy is used to optimise the particle-bond system parameters through evolving parameter sets. We tested our method both on artificially generated data and on data collected from real-life motion.

Results and comparisons between our method and other approaches show the advantage of using active particle-bond systems for image animation applications.

Keywords: computer animation, computer graphics, physically based motion modelling, particle-based modelling, evolutionary strategies, motion analysis, neural networks, neural controllers