We propose a simple but efficient method to extract rules from the radial basis function (RBF) neural network. Firstly, the data are classified by an RBF classifier. During training the RBF network, we allow for large overlaps between clusters corresponding to the same class to reduce the number of hidden neurons while maintaining classification accuracy. Secondly, centers of the kernel functions are used as initial conditions when searching for rule premises by gradient descent. Thirdly, redundant rules and unimportant features are removed based on the rule tuning results. Simulations show that our approach results in accurate and concise rules.
展开▼
