Evolving Hybrid Neural Fuzzy Network for System Modeling and Time Series Forecasting

Abstract

This paper introduces an evolving hybrid fuzzy neural network-based modeling approach using neurons based on uninorms and sigmoidal activation functions in a feed forward structure. The evolving neural network simultaneously adapts its structure and updates its weights using a stream of data. Currently, learning from data streams is a challenging and important issue because often traditional learning methods are impracticable to handle nonstationary and dynamic environments from where data come from. Uninorm-based neurons generalize fuzzy neurons models based on triangular norms and co norms. Uninorms increase the flexibility and generality of fuzzy neurons because they can modify their processing capabilities by adjusting their identity elements. In addition to structural plasticity induced by evolving network structures, identity elements adjustment adds functional plasticity in neural network processing. A recursive procedure to granulate the input space and uncover the evolving neural network structure, and an extreme learning-based algorithm to learn network weights are developed to train the neural network. Computational results show that the evolving neural fuzzy network is competitive when compared with representative methods of the current state of the art in evolving modeling.