Abstract
A characteristic feature of the neural network models is the large number of parameters. A model offering many parameters usually gives rise to problems, and the variance contribution to the modeling error might be very high. Therefore, it is crucial to find the model with the optimal number of parameters. In this paper two techniques of selection of the optimal number of model parameters are described and compared: explicit and implicit regularization techniques. Model validation forms the final stage of an identification procedure with the aim of assessing objectively whether the identified model agrees sufficiently well with the observed data. In this paper the reliability of the correlation-based validation tests and the χ2-test is analyzed.
Zusammenfassung
Große Parameteranzahl ist ein typisches Merkmal von Neuromodellen. Ein Neuromodell mit großer Parameteranzahl ist gewöhnlich mit vielen Problemen belastet weil in diesem Fall der Einfluss der Varianz auf den Modellfehler erheblich ansteigt. Deshalb ist es entscheidend, ein Neuromodell mit optimaler Parameteranzahl zu erstellen. In dem vorliegenden Beitrag werden zwei Techniken für die Auswahl der optimalen Modellparameteranzahl untersucht und verglichen: explizite und implizite Regularisationstechniken. Die Zuverlässigkeitsprüfung des Modells bildet den letzten Schritt eines Identifikationsverfahrens, dessen Ziel es ist, die Übereinstimmung des identifizierten Modells mit den durch Beobachtung gewonnenen Daten objektiv zu beurteilen. In diesem Beitrag wird die Zuverlässigkeit von Korrelationstests sowie des χ2-Tests analysiert.
Similar content being viewed by others
References
Hornik, K., Stinchcombe, M., White, H.: Multilayer feedforward networks are universal approximators. Neural Networks, Vol. 2 (1989), pp. 359–366.
Fahlman, S. E., Lebiere, C.: The cascade-correlation learning architecture. Advances in Neural Information Processing Systems 2 (1990), pp. 524–532.
Petrović, I., Baotić, M., Perić, N.: A cascade-correlation learning network with smoothing. Proc. of ICSC/IFAC NC ’98 Symposium (1998), pp. 1023–1029.
Zhao, Y., Atkeson, C. G.: Implementing projection pursuit learning. IEEE Transaction on neural networks. Vol. 7, No. 2 (1996), pp.362–373.
Reed, R.: Pruning algorithms — a survey. IEEE Transactions on neural networks, Vol. 4, No. 5 (1993), pp. 740–747.
Moody, J.: The effective number of parameters: An analysis of generalization and regularization in nonlinear learning systems. Advances in Neural Information Processing Systems 4 (1992).
Sjöberg, J., Ljung, L.: Overtraining, regularization, and searching for minimum in neural networks. Proc. of 4th IFAC Symposium on Adaptive Systems and Signal Processing (1992), pp. 669–674.
Leontaritis, I. J., Billings, S. A.: Input-output parameteric models for non-linear systems. Part I: Deterministic non-linear systems. International Journal of Control, Vol. 41, No. 2 (1985), pp. 303–328.
Chen, S., Billings, S. A., Grant, P. M.: Non-linear system identification using neural networks. International Journal of Control, Vol. 51, No. 6 (1990), pp. 1191–1214.
Sjöberg, J., Zhang, Q., Ljung, L., Benveniste, A., Deylon, B., Glorennec, P. Y., Hjalmarsson, H., Juditsky, A.: Non-linear black-box modeling in system identification: a unified overview. Automatica, Vol. 31, No. 12 (1995), pp. 1691–1724.
Bazaraa, S. M., Sherali, D. H., Shety, C. M.: Nonlinear programming — theory and algorithms. New York: John Wiley & Sons, 1993.
Ljung, L.: System identification: Theory for the user. Englewood Cliffs, New Jersey: Prentice-Hall, 1987.
Billings, S. A., Voon, W. S. F.: Correlation based model validity tests for non-linear models. International Journal of Control, Vol. 44, No. 1 (1986), pp. 235–244.
Leontaritis, I. J., Billings, S. A.: Model selection and validation methods for non-linear systems. International Journal of Control, Vol. 45, No. 1 (1987), pp. 311–341.
Petrović, I., Baotić, M., Perić, N.: An efficient Newton-type learning algorithm for MLP neural networks. Proc. of ICSC/IFAC NC ’98 Symposium (1998), pp. 551–557.
Nguyen, D., Widrow, B.: Improving the learning speed of 2-layer neural networks by choosing initial values of the adaptive weights. International Joint Conference on Neural Networks, Vol. 3 (1990), pp. 21–26.
Author information
Authors and Affiliations
Rights and permissions
About this article
Cite this article
Petrović, I., Baotić, M. & Perić, N. Regularization and validation of neural network models of nonlinear systems. Elektrotech. Inftech. 117, 24–31 (2000). https://doi.org/10.1007/BF03161395
Published:
Issue Date:
DOI: https://doi.org/10.1007/BF03161395