We introduce Recursive Exponential Mixed Models (REMMs) and derive the gradient of the parameters for the incomplete-data likelihood. We demonstrate how one can use probabilistic inference in Conditional Gaussian (CG) graphical models, a special case of REMMs, to compute the gradient for a CG model. We also demonstrate that this approach can yield simple and effective algorithms for computing the gradient for models with tied parameters and illustrate this approach on stochastic ARMA models.

PDF file

In: Proceedings of Tenth International Workshop on Artificial Intelligence and Statistics

Publisher: The Society for Artificial Intelligence and Statistics
Copyright © 2005 by The Society for Artificial Intelligence and Statistics.

Details