Linear and quadratic local models for ICE-Networks
M. Schafer, Werner Dilger · 2004
ICE-Networks are hybrid Neural Networks that are capable of fast initial learning and continuous learning, They have been developed for predicting the development of technical processes. ICE-Networks have a dynamic structure, they are built up starting from empty networks during the training process. This construction process is continued as long as the network is in use thus the network can yield an actual prognosis at any time. An ICE-Network is a layered network consisting of four layers. The units of the first hidden layer are RBF-neurons, called prototypes, and combine subsets of input vectors into so called local models that are maintained in the units of the second hidden layer. The type of the local models can be predefined by the developer of the ICE-Network, they can be linear or of higher order. In this paper the preciseness of the prognosis made by linear and quadratic models and the efficiency of computing those models are compared.