Multiobjective design of evolutionary hybrid neural networks

The paper presents a new approach to data-driven modeling. The models are flexibly configured in compliance with the neural network formalism, by accepting partially interconnected structures and various types of global and local neurons within each hidden neural layer. A simultaneous selection of convenient model structure and parameters is performed, making use of multiobjective graph genetic programming. For an efficient assessment of individuals, the authors suggest a new Pareto-ranking strategy, which permits a progressive combination between search and decision, tailored to handle objectives of different priorities. The experiments carried out for the identification of an industrial system show the capacity of the proposed approach to automatically build simple and precise models, whilst dealing with noisy data and poor aprioric information.