engleski

Convex Optimization in Training of CMAC Neural Networks

Simplicity of structure and learning algorithm plays important role in a real-time application of neural networks. The Cerebellar Model Articulation Controller (CMAC) neural network, with associative memory type of organization and Hebbian learning rule, satisfies these two conditions. But, Hebbian rule gives poor performance during on-line identification, which is used as a preparation phase for on-line implementation. In this paper we show that optimal CMAC network parameters can be found via convex optimization techniques. For standard l-2 approximation this is equivalent to the solution of Quadratic Program (QP), while for l-1 or l-inf approximation it is enough to solve Linear Program (LP). In both cases physical constraints on parameter values can be included in an easy and straightforward way.

CMAC neural networks; convex optimization; identification; control

nije evidentirano