A global learning method of RBFN

Radial basis function networks have been used successfully in various fields. Since the methods of learning RBFN are often separated into two stages which trend lead to suboptimal results. We proposed the method of using the EM algorithm to training the whole parameters at the same stage such that the parameters are learned globally. The initial parameters are decided by an improved cluster method to alleviate the local minimal problem. We analyze the relationship between the RBFN and the Gaussian mixture model that assure the feasibility of using the EM algorithm in RBFN.