Modular neurocontrollers for reaching movements

Abstract

Robotic controllers take advantage from neural network's learning capabilities as long as the dimensionality of the problem is kept moderate. This paper illustrates the possibilities offered by the combination of several neural networks to design more complex modular controllers. We propose a bi-directional architecture to derive the learning rules of the modules. The neurocontroller is trained globally, based on the interactions of the system with its environment, as one would do for a single network. The approach is evaluated on a robotic reaching application. The modular decomposition does not affect the controller interface. The computational cost is reduced and the rapid and efficient learning is maintained.