Real-time learning of aircraft parameters using recursive least-squares to train rbf networks

摘要

We describe a real-time algorithm for learning aircraft parameters to be used by an adaptive controller. Learning consists of training a collection of radial basis function neural networks to approximate the incoming data stream in the least-squares sense. Only the heights of the basis funtions are trained; heuristics are used to find their centers and widths. Since the heights enter the equations linearly, we employ recursive least-squares to quickly obtain the new heights when we incorporate additional data points and basis functions. In order to keep the computations manageable, we break the data stream into segments, with each segment approximated by about 10 basis functions. We illustrate the algorithm on a set of F-15 flight data.