A Possibilistic Approach to RBFN Centers Initialization

Clustering techniques have always been oriented to solve classification and pattern recognition problems. This clustering techniques have been used also to initialize the centers of the Radial Basis Function (RBF) when designing an RBF Neural Network (RBFNN) that approximates a function. Since classification and function approximation problems are quite different, it is necessary to design a new clustering technique specialized in the problem of function approximation. In this paper, a new clustering technique it is proposed to make the right initialization of the centers of the RBFs. The novelty of the algorithm is the employment of a possibilistic partition of the data, rather than a hard or fuzzy partition as it is commonly used in clustering algorithms. The use of this kind of partition with the addition of several components to use the information provided by the output, allow the new algorithm to provide better results and be more robust than the other clustering algorithms even if noise exits in the input data.