Abstract
Radial basis function (RBF) networks are often used for identification of nonlinear dynamic systems. The main reason why RBF networks are so successful is that the hidden layer parameters can be fixed in a very reasonable way and only the weights are optimized by a standard least squares technique. Thus for the case of Gaussian RBFs a good choice of the centers and standard deviations is crucial for good network performance. We show that Ihe most widely used clustering approach has many drawbacks. An alternative technique for center determination is presented, that is not completely unsupervised but exploits error information. It is based on a fusion of linear parameter estimation and the RBF network. First, a linear system is estimated from data. Then only the nonlinear part is approximated by an RBF network.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have