A self-organizing neuro-fuzzy network based on first order effect sensitivity analysis

As an effective method that can provide the information about the influence of inputs on the variation of output, variance based sensitivity analysis is widely used to determine the structure of neural networks. In the past, the global sensitivity analysis method for the total effect has been used for the structure learning of neural networks and various growing and pruning algorithms have been developed. In this paper, we find that neuro-fuzzy networks have the characteristics of additive models in which the first order effect index of the influence can provide the same comprehensive information as the total effect index, thus we only need to analyze the first order effects of the inputs to their output layers. Based on this observation, many low-cost effective methods for the first order effect global sensitivity can be used in for developing self-organizing neuro-fuzzy networks. Specifically, Random Balance Designs is employed here for sensitivity analysis. In addition, we also introduce the concept of systemic fluctuation of neuro-fuzzy networks to determine whether adjustment is needed for a network. This concept helps us to build a new procedure about the leaning of self-organizing neuro-fuzzy networks and to accelerate its speed of convergence in learning and organizing. Examples of simulations have demonstrated that our proposed method performs better than other existing procedures for self-organizing neuro-fuzzy networks, especially in learning of the network structure. (C) 2013 Elsevier B.V. All rights reserved.