Factor Sensitivity Analysis for Multivariable Systems Using Bayesian Neural Networks

Neural interpretation is of increasing interest in artificial neural networks and it is potential to reveal the intrinsic mechanism of multivariable systems. This study aims at investigating the efficiency of Bayesian neural networks in neural interpretation. The measures to ensure the stability of the network model are first elaborated and then, two types of Bayesian networks with linear and partly-linear transfer functions are exploited to conduct neural interpretation for simulated multivariable systems. Experimental results show that aided with connection weight calculation, Bayesian neural networks own distinct advantages over other network models in evaluating the relative contribution of independent variables to single dependent one. Therefore, the method proposed in this study is promising to perform factor sensitivity analysis for multivariable systems.