Multiobjective genetic programming with adaptive clustering

This paper presents a new approach meant to provide an automatic design of feed forward neural models by means of multiobjective graph genetic programming. The suggested algorithm can deal with partially interconnected neural architectures and various types of global and local neurons within each hidden neural layer. It concomitantly ensures the reduction of variables and the selection of convenient model structures and parameters, by working on a set of graph-based encrypted individuals built via genetic programming with the guarantee of phenotypic and genotypic validity. In order to provide a realistic assessment of the neural models, the optimization is carried out subject to multiple objectives of different priorities. In relation to this idea, the authors propose a new Pareto-ranking strategy, which progressively guides the search towards the preferred zones of the exploration space. The fitness assignment procedure monitors the phenotypic diversity of the best individuals, as well as the convergence speed of the algorithm, and exploits the resulted heuristics for performing a preliminary clustering of individuals. The experimental trials targeting the identification of an industrial system show the capacity of the suggested approach to automatically build simple and precise models, whilst dealing with noisy data and scarce a priori information.