ABSTRACT
Most learning algorithms of artificial neural networks are dealing with the efficiency for the teaching patterns only. However, when the trained network is applied to the specific task, the network performance associated with the generalization capacity including the untrained data becomes more important. From this point of view, we consider a new method of on-line learning of a network with Gaussian kernel functions. The suggested approach in this paper has two learning phases, the initial off-line learning phase for the given teaching patterns and the on-line learning phase for the incoming data. At the initial phase of learning, we estimate the necessary number of kernel functions as well as the parameters associated with kernel functions for the given teaching patterns. This process is assumed to be done by the batch process (or off-line learning). After this initial set-up of network structure, on-line learning for the incoming data is done by recruiting the necessary number of kernel functions when the incoming data need to be accommodated to maintain or improve the network performance. To show the effectiveness of our approach, simulation results for the prediction of the Mackey-Glass chaotic time series are presented.
