Studies into computational intelligence approaches for the identification of complex nonlinear systems

2016-04-26This study builds on major advances in the field of Computational Intelligence to develop a state‐of‐the‐art data‐driven methodology that provides parsimonious optimized computational models in the form of systems of differential equations that characterize the behavior of complex nonlinear phenomena observed in mechanical and biological systems. The proposed hybrid identification scheme integrates various stochastic optimization methods and computer algebra techniques, such as Genetic Programming and Genetic Algorithms, to evolve structures of differential equations, to optimize their parameters, and to reduce their complexity for controlling bloat. The investigated scenarios include systems that exhibit polynomial‐type nonlinearities in their response, systems that show discontinuity in their nonlinear behavior, systems with memory‐dependent and dissipative characteristics, as well as the human spine. The investigations are conducted by processing input and output data obtained from synthetic simulations as well as experiments. It is shown that the proposed technique yields reduced‐order, reduced‐complexity, optimized differential equations, that accurately characterize the behavior of the investigated systems, and provide accurate estimates. The generalization extent of the discovered models is scrutinized by assessing their performance in new dynamical environments through applying validation excitations that are substantially different from the excitations employed for training. Findings reveal that the resulting models provide reasonably accurate estimates, even when models are subjected to new stimulations with various intensities. Thus, the proposed approach of this study presents a robust data‐driven methodology based on evolutionary computation techniques that provides elegant computational models to represent variety of complex nonlinear systems