Error Bounds for Steel Tensile Strength Prediction—Comparison of Bayesian and Ensemble Model Approaches

Abstract

One of the obstacles to applying neural networks in industry is a lack of confidence in model predictions. Recently, there has been increased interest in tackling this lack of confidence by adding error bounds on predictions. In this paper, Bayesian networks and ensemble modelling approaches have been applied to establish the error bounds. A brief introduction to these two approaches is given together with discussion on their merits and drawbacks for error bounds calculation. The two error bound calculation methods are demonstrated with modelling examples from simple function to complicated industrial steel tensile strength data. We show that for complicated problems, creditable error bounds are very difficult to obtain via the Bayesian approach, since it involves many design parameters which are very difficult to define and because some assumptions made in the Bayesian approach on error bound calculation are hard to satisfy. The ensemble modelling approach, however, gives a much robust prediction on error bounds. It shows that the error bounds for the steel tensile strength prediction are consistent with existing metallurgical experience.