An improved novelty criterion for resource allocating networks

The author introduces a new novelty criterion for resource allocating RBF networks (RANs) based on standard signal processing theory. This network growth prescription is considerably less sensitive to noise and outliers than those of previous RANs, and also removes the need for ad-hoc hyperparameters. An added advantage of this novelty criterion is that, as it is independent of the parameters of the extended Kalman filter training algorithm, the filter can be modified for application to slowly varying nonstationary environments without adversely affecting the network's capacity for growth. The author demonstrates the relative improvement of this criterion on two non-stationary real-world problems: electricity load forecasting and exchange rate prediction.