"Towards Robust Evolving Fuzzy Systems"
, in Plamen Angelov, Dimitar Filev, Nikola Kasabov: Evolving Intelligent Systems: Methodologies and Applications, John Wiley & Sons, Seite(n) 87-126, 4-2010
Towards Robust Evolving Fuzzy Systems
Sprache des Titels:
Evolving Intelligent Systems: Methodologies and Applications
In this chapter, methodologies for more robustness and transparency in
evolving fuzzy systems will be demonstrated. After outlining the evolving
fuzzy modelling approaches FLEXFIS (=FLEXible Fuzzy Inference Systems) for fuzzy regression models
and FLEXFIS-Class (=FLEXible Fuzzy Inference Systems for Classification) for fuzzy classification models,
robustness improvement strategies during the incremental learning phase will be explained, including
regularization issues for overcoming instabilities in the parameter estimation process,
overcoming the unlearning effect, dealing with drifts in data streams, ensuring a better extrapolation behavior
and adaptive local error bars serving as local confidence regions surrounding the evolved fuzzy models.
The chapter will be concluded with evaluation results on high-dimensional real-world data sets, where
the impact of the new methodologies will be presented.