Online Learning of Data Streams: Evolving Fuzzy Predictor with Multi- variable Gaussian Participatory Learning and Recursive Weighted Total Least Squares

Resumo

This paper introduces an evolving fuzzy system called eFTLS (evolving Fuzzy with Multivariable Gaus-
sian Participatory Learning and Recursive Weighted Total Least Squares) constructed based on a non-supervised

recursive clustering algorithm with participatory learning and multivariate Gaussian membership functions. The

eFTLS uses a clustering algorithm to extract the first-order Takagi-Sugeno functional rules. The clustering algo-
rithm can add a new cluster, delete, merge, or update existing clusters. The clusters are created using a compatibility

measure and an alert mechanism. The compatibility measure is computed by Euclidian or Mahalanobis distance

according to the number of samples in the cluster. An age and population based-method excludes inactive clus-
ters. Redundant clusters are merged whenever there is a noticeable overlap between two clusters. An algorithm

of recursive weighted total least squares updates the consequent parameters. The performance of the eFTLS is
evaluated and compared with alternative state-of-the-art models in forecasting tasks. Computational experiments
and comparisons suggest that the eFTLS perform better or are similar to alternative models.