An improved algorithm for online identification of evolving TS fuzzy models

In this paper an improved algorithm for online identification of Takagi-Sugeno fuzzy rule-based models from I/O data streams is proposed. The TS model has evolving structure i.e the fuzzy rules can be added, modified or deleted automatically. Both parts of identification algorithm (unsupervised fuzzy rule-base antecedent learning by a recursive, non-iterative clustering, and the supervised linear sub-model parameters learning by RLS estimation) are developed for the MIMO case. The radius of influence of each fuzzy rule is calculated as an adaptive vector instead of being fixed vector, allowing different areas of data space to be covered. The centers and widths of membership functions initially determined by online clustering are optimized continuously using a gradient descent method. This feature enables the identification algorithm to deal with time-varying systems and non-stationary data streams. Simulation studies (using two benchmark problems) and comparisons with some other online learning algorithms demonstrate that a more compact structure with higher performance is achieved by the proposed approach.