Physical Models That Learn

In this paper is proposed an analysis method for Physical Model networks, based on connectionist learning algorithms. The first experimental results, obtained on limited cases, are quite encouraging and suggest some improvements, using the interactivity of Physical Models. 1 From synthesis to analysis Synthesis and analysis form a pair of two symmetrical processes: while synthesis turns computer parameters into sound, analysis attemps to infer, from a sound behaviour, the parameters of the model in the machine. Given a physically based sound synthesis system, this paper deals with analysis techniques that have to be associated with it. CORDIS [Cadoz et al., 1993] is a physical description language for music instruments. It has specificities that generate some constraints in the corresponding analysis system: - instead of considering large differential equations from the instrument to be modeled and then numerically integrating them, it manipulates a large number of simple physical auto..