Learning data streams online - An evolving fuzzy system approach with self-learning/adaptive thresholds

Recognizing the weakness of the existing evolving fuzzy systems (EFSs) where the selection and determination of thresholds for the structure and parameter learning are completely relied on the trial and error strategy, this paper proposes a novel and symmetrical approach with self-learning/adaptive thresholds (EFS-SLAT) for EFSs to overcome such a fundamental weakness. Departing from the common but implicit assumption in the existing EFS approaches that the thresholds are fixed parameters throughout the learning process, EFS-SLAT treats the thresholds as dynamical parameters which are varying with the evolution of systems being learned. By utilizing the online training errors as an indicator to reflect the underfitting and overfitting state of an EFS model, the proposed EFS-SLAT selects and adjusts the values of threshold parameters automatically and dynamically based on the evolving speed and nonlinear degree of the EFS. By testing EFS-SLAT on several well-known benchmark examples, and comparing it with many state-of-the-art approaches, EFS-SLAT is verified to be capable of giving preferable results. (C) 2019 Elsevier Inc. All rights reserved.