Radial basis function neural networks provide an attractive model for high dimensional nonparametric estimation, particularly for model-based process control. They are faster to train than conventional feedforward networks with sigmoidal activation functions , and provide a model structure better suited for adaptive control. Radial basis function neural networks can be viewed as a mixture model where the number of components is chosen adaptively. This view allows for statistical analysis and the extension of current radial basis function networks. In particular, this research will develop models using combinations of local linear models, invert the models for calibration, estimate the local prediction errors, identify outliers or novel data points, adaptively choose the number of basis functions and derive algorithms that learn incrementally. The research extends and improves a certain class of artificial neural networks, the radial basis function networks, for applications to (chemical) process control. Algorithms developed will be written so they can be efficiently updated for use in on-line process control applications. This is a coordinated project between two investigators, a statistician and a chemical engineer. A major chemical company, DuPont, has expressed strong interest in using these methods and they have supplied data to test the methodology of this research .
