I have been developing a hierarchical model that fundamentally consists of several equations, each of which is expressed as a function of summed mixes of other probabilistic components. Each summed component to date has been modeled as either a Gaussian
or the product of two or more Gaussians. This model does not work under EP but appears to work well under VMP and Gibbs.
One of the newer summed components I wish to add is a one-tailed continuous variable that is most simply modeled as either an exponential or heavy-tailed power/Pareto distribution. I had started by modeling this component as a Gamma with Shape parameter
of 1 to simulate an exponential distribution. This addition causes the model to fail to compile because the engine does not support Factor.Difference. Messages displayed are either "Gamma is not of Type Gaussian for argument 2 of Method DoubleMinusVmpOp...
under VMP and "double is not of type WrappedGaussian..." under Gibbs.
Do you have any suggestions as to how to best address this issue in Infer.net?