A new convex hull, sliding window based online adaptation method for fixed-structure radial basis function neural networks

Abstract

In any online adaptation scheme, two important phenomena should be taken into consideration; parameter shadowing and parameter interference. To alleviate these problems, in this paper a sliding window based online adaptation method for fixed-structure Radial Basis Function Neural Networks (RBFNNs) is proposed. The method is capable of updating the underlying model using the new arriving samples reflecting, to a good extent, new regions in the input-output space and also can deal with the two phenomena mentioned above. The online adaptation process requires a small update rate, while maintaining a good level of accuracy of the updated model.