Improving on-line neural networks backpropagation convergence speed with mixed pattern-batch learning

The present paper describes an algorithmic technique to speed up weight convergence in neural networks on-line training. Standard pattern backpropagation is modified to train the neural network over a time window of samples and not one sample only, so that a faster weight convergence may be achieved. The use of such training technique is explained in an adaptive control task and problems related to validation of real functional approximation are investigated.