Levenberg-Marquardt training for modular networks

The modular neural network has been shown to be an effective alternative to multilayer feedforward networks, especially for implementing functions with sharp changes. This paper describes a new method for training modular networks, based on the Levenberg-Marquardt algorithm for nonlinear least squares. The algorithm is tested on several function approximation problems, and the performance is compared with standard steepest ascent and the Rprop algorithm.