Determining the influence of input parameters on BP neural network output error using sensitivity analysis

Abstract

Neural networks, due to their excellent capabilities for modelling process behaviour, are gaining precedence over traditional empirical modelling techniques, such as statistical methods. While neural networks have good ability to map any reasonable continuous function, they do not easily explain how the inputs are related to an output, and also whether the selected inputs have any significant relationship with an output. There is quite often a need to identify some order of influence of the input variables on the output variable. In this paper, a technique for determining the order of influence of the n elements of the input vector on the m elements of the output vector is presented and discussed. While a sample mathematical function is used to introduce the technique, a more practical application of this method in the aluminium smelting industry is considered. It is shown that, using a sensitivity analysis on the backpropagation (BP) algorithm, the degree of influence of the input parameters on the output error can be successfully estimated.