|
Vol: 55(69) No: 4 / December 2010 Evolutionary Neural Networks with Heterogenous Hidden Neurons Lavinia Ferariu Department of Automatic Control and Applied Informatics, “Gheorghe Asachi” Technical University of Iasi, Faculty of Automatic Control and Computer Engineering, Bd. D. Mangeron 27, 700050 Iasi, Romania, phone: (0040) 232-278680, e-mail: lferaru@tuiasi.ro Bogdan Burlacu Department of Automatic Control and Applied Informatics, “Gheorghe Asachi” Technical University of Iasi, Faculty of Automatic Control and Computer Engineering , Bd. D. Mangeron 27, 700050 Iasi, Romania, e-mail: bburlacu@tuiasi.ro Keywords: genetic programming, directed acyclic graphs, neural networks, system identification Abstract The paper discusses the main challenges of nonlinear system identification, focusing on the benefits and the limitations of neuro-genetic approaches. Materials and methods: Feed-forward neural networks with heterogeneous hidden layers are considered. A flexible and efficient selection of proper neural architecture and parameters is performed by means of genetic programming based on graph encoding. To ensure good exploration capabilities, the hierarchical models are modularly built, using customized sets of primitives which guarantee both genotypic and phenotypic validity. The generation of offspring is based on genetic crossovers and mutations compatible with the hierarchical architecture of the evolved individuals. In order to allow the rejection of overfitted models and an efficient bloat reduction, the problem is approached via a multiobjective optimization, which targets the selection of accurate and simple models, by means of linear aggregation. The design software is implemented using the C programming language, in terms of high scalability, reduced memory consumption and small execution time. Results: The experimental trials carried out on the identification of two industrial subsystems reveal the flexibility of the approach in building appropriate neural models of different levels of architecture complexity. The software features high stress performances, being suitable for managing long evolutionary loops and populations of numerous, complex models. Conclusions: The presented approach offers increased flexibility in solving complex nonlinear system identification problems, when scarce apriorical information is available, numerous models are to be designed and/or high performances of accuracy are requested. References [1] P. J. Flemming, R. C. Purshouse “Evolutionary Algorithms in Control Systems Engineering: A Survey”, Control Engineering Practice 10, 2002, 1223-1241. [2] M. Affenzeler, Winkler S., Wagner S., Beham A., Genetic Algorithms and Genetic Programming: Modern Concepts and Practical Applications (Numerical Insights), Boca Raton, FL, CRC Press, 2009, 157-207. [3] S. Haykin,. Neural Networks - A Comprehensive Foundation, McMillan College Publishing Company, New York, 2nd Edition, 1999. [4] J. Patra, Pal R., Chatterji, B., Panda G., “Identification of nonlinear dynamic systems using functional link artificial neural networks”, IEEE Transactions on System, Man and Cybernetics, Part B: Cybernetics, 29, 254–262, (1999). [5] B. Igennik, Tabib – Azar M., Le Clair S. R., “A net with complex weights”, IEEE Transactions on Neural Networks, 12, 236–249, 2001. [6] L. Ferariu, M. Voicu, “Nonlinear System Identification Based on Evolutionary Dynamic Neural Networks wih Hybrid Structure” in Proc. of IFAC Congress, Prague, Czech Republic, 2005. [7] L. Ferariu, B. Burlacu, Graph Genetic Programming for Hybrid Neural Networks Design, IEEE International Joint Conferences on Computational Cybernetics and Technical Informatics, Timisoara, Romania, 2010. [8] R. Poli., Langdon W. B., Mc Phee N. F., A Field Guide to Genetic Programming. Published via http://lulu.com (with contributions of J.R. Koza), [Online]. Available: http://www.gp-field-guide.org.uk, 2008. [9] J. R. Koza, Genetic Programming: On the Programming of Computers by Means of Natural Selection. Cambridge, MA: MIT Press, 1992. [10] A Walker, Miller, J. F.. The Automatic Acquisition, Evolution and Reuse of Modules in Cartesian Genetic Programming. IEEE Transactions on Evolutionary Computation, 12 (4), 2008, 397-417. [11] G. N Beligiannis, Skarlas, L. V., Likothanassis, S. D., Perdikouri, K. G. Nonlinear model structure identification of complex biomedical data using a genetic-programming-based technique, IEEE Transactions on Instrumentation and Measurement, 54 (6), 2005, 2184-2190. [12] X.-T. Feng, Chen, B.-R., Yang, C., Zhou, H., Ding, X.. Identification of visco-elastic models for rocks using genetic programming coupled with the modified particle swarm optimization algorithm, International Journal of Rock Mechanics and Mining Sciences,43 (5), 2006, 789-801. [13] J. U Dolinsky, Jenkinson, I. D., Colquhoun, G. J.. Application of genetic programming to the calibration of industrial robots, Computers in Industry, 58 (3), 2007, 255-264. [14] T Marcu., Mirea L., Ferariu L., Frank P. M. “Miscellaneous Neural Networks Applied to Fault Detection and Isolation of an Evaporation Station”, Proc. of 4th IFAC Symposium on Fault Detection, Supervision and Safety for Technical Processes, SAFEPROCESS , Hungary, 2000. |