|
"
Apprentissage dans les réseaux récurrents pour la modélisation mécanique et étude de leurs interactions avec l'environnement
"
Download (in french): |
Recurrent neural networks are used to model complex dynamic behaviours and to reproduce - to learn - such behaviours. These adaptive properties can be applied to physical modelling networks dedicated to the simulation of musical instruments.
The parameters of physical modelling nets (inertia, stiffness and damping) can be learned in order to reproduce a given mechanical behaviour thanks to recurrent neural networks. Classical learning algorithms as well as original ones are developed and simulated. Physical networks tackle the issue of the interaction with the environment. Thanks to several experiments, it is shown that, under certain conditions, the interaction makes the learning succeed, especially if the interaction with the environment allows learning of progressive complexity. Finally, an original system for analysis of musical instruments is proposed, which involves the real-time interaction of an instrumentalist, an instrument and a computer that simulates the adaptive network. |
(( TOPICS
Physical modeling Progressive learning (( LABS Lifia Acroe |