Decremental Learning of Evolving Fuzzy Inference Systems: Application to Handwritten Gesture Recognition

This paper tackles the problem of incremental and decremental learning of an evolving and customizable fuzzy inference system for classi cation. We explain the interest of integrating a forgetting capacity in such an evolving system to improve its performances in changing environment. In this paper, we describe two decremental learning strategies, to introduce a forgetting capacity in evolving fuzzy inference systems. Both techniques use a sliding window to introduce forgetting in the optimization process of fuzzy rules conclusions. The rst approach is based on a downdating technique of least squares solutions for unlearning old data. The second integrates di ered directional forgetting in the covariance matrices used in the recursive least square algorithm. These techniques are rst evaluated on handwritten gesture recognition tasks in changing environments. They are also evaluated on some wellknown classi cation benchmarks. In particular, it is shown that decremental learning allow to adapt to concept drifts. It is also demonstrated that decremental learning is necessary to maintain the system capacity of learning new classes over time, making decremental learning essential for the life-time use of an evolving and customizable classi cation system.