A Constructive-Fuzzy System Modeling for Time Series Forecasting

This paper suggests a constructive fuzzy system modeling for time series prediction. The model proposed is based on Takagi-Sugeno system and it comprises two phases. First, a fuzzy rule base structure is initialized and adjusted via the expectation maximization optimization technique (EM). In the second phase the initial system is modified and the structure is determined in a constructive fashion. This phase implements a constructive version of the EM algorithm, as well as adding and pruning operators. The constructive learning process reduces model complexity and defines automatically the structure of the system, providing an efficient time series model. The performance of the proposed model is verified for two series of the reduced data set at the Neural Forecasting Competition, for one to eighteen steps ahead forecasting. Results show the effectiveness of the constructive time series model.