Performance of the RBF networks (RBFNs) depends on numerous factors. One of the most significant in this respect is the network structure. Designing an effective structure is the task carried-out at the network initialization stage where the number of centroids with their respective locations need to be calculated or induced. It is well known that setting-up values of these parameters is NP-hard. Usual approach to deal with the problem is to decide on the number of hidden units and to apply the k-means algorithm to calculate cluster centroids. Unfortunately, RBFNs designed in such a conventional way may not be accurate, as the number of clusters at the initialization stage must be set up a priori. To overcome the problem we have proposed the similarity-based algorithm for the RBF network initialization. In this paper the approach is extended through proposing an alternative method to initialize RBFNs using the kernel-based fuzzy clustering algorithm. In both cases the number of resulting centroids and their initial locations are provided by the respective algorithm. A comparative study of both approaches to RBFNs initialization is included and their effectiveness is demonstrated on artificial and real datasets.

Login to Download PDF

Please login with your Paper ID and password to access the full PDF.

🔑 Login to Download