ABSTRACT
We modify the mixtures-of-experts model by utilizing a different parametric form for the gating network. The modified model is trained by an EM algorithm. In comparison with earlier models—trained by either EM or gradient ascent—there is no need to select a learning stepsize to guarantee the convergence of the learning procedure. We report simulation experiments that show that the new architecture yields significantly faster convergence. We also utilize the new model to perform piecewise nonlinear function approximation with polynomial, trigonometric, or other prespecified basis functions.
