New Technique for Initialization of Centres in TSK Clustering-Based Fuzzy Systems

Several methodologies for function approximation using TSK systems make use of clustering techniques to place the rules in the input space. Nevertheless classical clustering algorithms are more related to unsupervised learning and thus the output of the training data is not taken into account or, simply the characteristics of the function approximation problem are not considered. In this paper we propose a new approach for the initialization of centres in clustering-based TSK systems for function approximation that takes into account the expected output error distribution in the input space to place the fuzzy system rule centres. The convenience of proposed the algorithm comparing to other input clustering and input/output clustering techniques is shown through a significant example.