Adaptive Modularity and Time-Series Prediction

  • Mira Trebar
  • Andrej Dobnikar


This paper focuses on the use of recorded time-series to estimate future values as a function of their past values. We study the local events in input space and apply them as classes of similar patterns to the problem of short-term prediction. The decomposition of the time-series into the patterns formed from d past values denoted as an input vector and the true future value in the observed time-series is performed. From the observation of past values we can conclude that similar input vectors often have similar predictive values. We assume that this principle can be expanded in the future. The time-series is based on a similarity measure partitioned into similar patterns grouped into classes. Each of these classes computes the predicted value. The final predicted value is then determined with only one class obtained by the classification from the present input vector.

We apply the concept of modularity in neural networks to perform the local computation of predicted values based on classes of similar patterns. The approach under discussion is a combination of classification and prediction problems. The classification of the input space defines the classes of similar patterns. The number of classes is adaptive and based on the length of the observation pattern for time-series. The predicted value in time-series is obtained by the class modular neural network, which generalises from the classes of similar patterns in training sets in combinations of classification input data into the class of similar patterns.


Neural Network Input Vector Input Pattern Prediction Module Class Neural Network 
These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.


Unable to display preview. Download preview PDF.

Unable to display preview. Download preview PDF.


  1. [1]
    Haykin, S.: Neural Networks — A Comprehensive Foundation, Macmillan College Publishing Company, New York 1994MATHGoogle Scholar
  2. [2]
    Jacobs, R.A., Jordan, M.L: Learning Piecewise Control Strategies in a Modular Neural Network Architecture, IEEE Trans, on Syst., Man, and Cyb., Vol.23, No.2, pp. 337–345, 1993MATHCrossRefGoogle Scholar
  3. [3]
    Ronco, E., Golle, H., Gawthrop, P.: Modular Neural Networks and Self-Decomposition, Technical Report: CSC-96012, February 11, 1997Google Scholar
  4. [4]
    Petridis, V., Kehagias, A.: Modular Neural Network for MAP Classification of Time-series and the Partition Algorithm, IEEE Trans, on Neural Networks, Vol. 7, No. 1, pp. 73–86, January 1996CrossRefGoogle Scholar
  5. [5]
    Pham, D.T., Xing, L.: Neural Networks for Identification, Prediction and Control, Springer Verlag, London, 1995CrossRefGoogle Scholar
  6. [6]
    Pole, A., West, M., Harrison, J.: Applied Bayesian Forecasting and Time Series Analysis, Chapman & Hall, New York, 1994MATHGoogle Scholar
  7. [7]
    Trebar, M., Dobnikar, A.: Time Series Prediction with Modular Neural Network Architecture, SOCO′97 Second International ICSC Symposium on Soft Computing, Nimes, France, pp. 152–155, 1996Google Scholar
  8. [8]
    Weigend, A.S., Huberman, B.A., Rumelhart, D.E.: Predicting the Future: A Connectionist Approach. International Journal of Neural Systems, Vol. 1, N°3, pp. 193–209, 1990CrossRefGoogle Scholar

Copyright information

© Springer-Verlag Wien 1999

Authors and Affiliations

  • Mira Trebar
    • 1
  • Andrej Dobnikar
    • 1
  1. 1.Faculty of Computer and Information ScienceUniversity of LjubljanaLjubljanaSlovenia

Personalised recommendations