Diseño de sistemas inteligentes en plataformas de cómputo paralelas

Resumen

The work carried out in this thesis deals with the problem of designing Radial Basis Function Neural Networks (RBFNNs) for function approximation or regression. The main innovations proposed in this thesis are: 1. The development of a new clustering algorithm that, improving the drawbacks presented in the CFA algorithm, is able to perform an adequate initialization of the centers of a RBFNN for function approximation problems. 2. The study of the different kinds of possibilistic clustering approaches when they are adapted to the function approximation problem. 3. A new algorithm, the OVI, that is able to perform an appropriate initialization of both the centers and the radii, obtaining better results than classical heuristics used for radii initialization, i.e., kNN and CVI. 4. A new methodology to design an RBFNN considering the optimization of the parameters of the RBFs, the structure of the network, and the input variables using a multiobjective approach. This new algorithm required the design of new crossover and mutation operators. 5. The parallelization of the algorithm through heterogeneous island specialization, where each island uses different crossover and mutation operators to evolve the population. 6. The integration of two different design methodologies into a single algorithm that takes advantage of the exploitation capabilities of local search algorithms to obtain adequate solutions after exploring the solution space through the global optimization algorithm. 7. A new interface that allows programs compiled in MATLAB to call MPI functions of any library that follows the MPI standard. 8. A new concept of fuzzy dominance that allows the algorithm to keep a balance between the size of the networks and the approximation accuracy. 9. The application of the proposed methodology to real world problems.