Overcome neural limitations for real world applications by providing confidence values for network prediction

In this paper we present an incremental construction algorithm for continuous learning tasks and one of its special features-simultaneous learning of the target function and a confidence value for the system predictions. The basis of the hybrid system is a radial basis function (RBF) network layer. The second layer consists of local models. The two layers are closely combined with a strong interaction. The number of RBF-neurons and the number of local models have not to be determined in advance. This is one of the main advantages of the algorithm. Another advantage emphasized in this paper is the ability to learn the training data distribution simultaneously to the learning of the target function. The learned data set distribution can be used as a confidence value for a given network prediction. The development of the described approach is embedded in a larger project that is primarily concerned with system identification tasks for industrial control such as steel processing.