Structure Identification for Complex System and Inference Interpretable Rules

Abstract

The large and complex systems are characterized by the difficulty of formalizing and the lack of expertise. They are built from data representing instances of the system. The drawback of these structure selection methods is that pay particular attention to the numerical accuracy of the resulting model and little attention to the qualitative and semantic aspect. The use of a large number of input variables results in an introduction of redundant elements, poor transparency and an excessive complexity of the model obtained. To solve these problems, a particular interest should be given for selecting relevant input variables that can provide a reasonable compromise between the quality of approximation, the complexity and the transparency of the model. The proposed approach for structure selection assumes constraints on the number of variables to be selected in the initial combination to describe the model. Relevant inputs are found from input-output data and variables with the highest correlation coefficient with the output variable are selected in the initial combination.