ABSTRACT
Recurrent or dynamical neural networks have been widely used in fields such as optimization, identification and controls, and also in the synthesis of associative memories: see [154, 156, 158, 163, 164, 166, 167, 169, 172, 173, 174, 177, 178, 179] and ref erences therein. On the other hand, wavelet-based networks have been recently introduced as an alternative to classical feed forward schemes in approximation, classification, detection, and nonparametrie estimation problems [160,161,183,184], Wavelet theory has also been used in the analysis and design of different feed-forward neural systems [168],
In this context, the purpose of the present work is to present an extended discussion of some qualitative properties and simu lation results concerning wavelet-based recurrent networks pre viously addressed in [171], Specifically, the use of a waveletbased selection of the activation function in additive networks
will be studied. The purpose of this approach is two fold: on the one hand, the oscillatory nature of wavelets allows for an easy design of networks with multiple equilibria. On the other hand, the sealing and translation parameters of wavelets provide some hints for an appropriate selection of initial values in the training process.
