Optimum back propagation network conditions with respect to computation time and output accuracy

Abstract

An important consideration when designing neural network training data is to carefully select those variables that are to be used as inputs. Only those parameters that contribute towards improving the accuracy of the network's prediction should be included as input parameters. Despite a large variety of neural network models, backpropagation (BP) is the most commonly applied model for an extensive range of applications. However, when applying BP networks to process modelling or control, it is necessary to select the correct network architecture and activation functions in order to minimise the computation time and maximise the network's accuracy. In addition, in order to improve network performance, it is necessary to use sufficient training data, spanning a comprehensive input range. While many of the techniques for improving network performance are based on a heuristic approach, some important aspects are detailed in this paper for selecting the optimum network conditions, with respect to computation time and accuracy, using a mathematical function as a sample application.