Enhanced algorithm for randomised model structure selection

Abstract

Model structure selection for nonlinear system identification has been widely studied over the last 30 years due to its great importance. There are many methods in the literature to deal with structure selection, although these methods have their specific benefits, they face some difficulties in selecting the structure for a parsimonious model. In this paper, two methods based on the Randomised Model Structure Selection (RaMSS) approach are introduced in order to deal with the structure selection problem. The first one is the Randomised Model Structure Selection with Error Reduction Ratio (RaMSS-ERR) that uses the error reduction ratio as a filter for the terms analysis, improving the convergence, and the second one is the Randomised Model Structure Selection with Genetic Inheritance (RaMSS-EGI) that uses a genetic inheritance in order to get a faster convergence. The methods were applied to benchmark models and the results are encouraging. Applications to systems with a large candidate regressor set and to a continuous stirred-tank reactor are also carried out. The results show that the proposed method may be used to identify both linear and nonlinear model structures with a reduced number of iterations, computational time, and number of explored models.