Evolving predictive neural models for complex processes

Matteo De Felice, Mauro Annunziato, Ilaria Bertini, Stefano Panzieri, Stefano Pizzuti

Research output: Contribution to conferencePaper

Abstract

In this work we present a study on Artificial Neural Networks (ANN) performance with different topologies and training algorithms in order to develop a model of a dynamical system, focusing on the effect of unknown inputs (disturbance). Therefore, we study the ANN performance in model identification and prediction by training them using traditional gradient based and evolutionary methods. Tests have been made with different prediction horizons on two experimentations: without disturbance and with a pulse train unknown disturbance. Results show that without disturbances the performance of gradient trained ANN is slightly better than that of evolutionary trained ANN. The situation is different when in the presence of disturbance: gradient trained ANN performance gets much worse compared to evolutionary ANN.

Conference

Conference4th International Conference on Cybernetics and Information Technologies, Systems and Applications, CITSA 2007, Jointly with the 5th International Conference on Computing, Communications and Control Technologies, CCCT 2007
CountryUnited States
Period1/1/07 → …

    Fingerprint

All Science Journal Classification (ASJC) codes

  • Computer Networks and Communications
  • Information Systems

Cite this

De Felice, M., Annunziato, M., Bertini, I., Panzieri, S., & Pizzuti, S. (2007). Evolving predictive neural models for complex processes. Paper presented at 4th International Conference on Cybernetics and Information Technologies, Systems and Applications, CITSA 2007, Jointly with the 5th International Conference on Computing, Communications and Control Technologies, CCCT 2007, United States.