One of the authors has recently implemented two new variable nodes Harva04DI, Harva05UAI into the Bayes Blocks software library. They are the mixture-of-Gaussians (MoG) node and the rectified Gaussian node. MoG can be used to model any sufficiently well behaving distribution Bishop95. In the independent factor analysis (IFA) method introduced in Attias99, a MoG distribution was used for the sources, resulting in a probabilistic version of independent component analysis (ICA) ICABook01.
The second new node type, the rectified Gaussian variable, was introduced in Miskin01RE. By omitting negative values and retaining only positive ones of a variable which is originally Gaussian distributed, this block allows modelling of variables having positive values only. Such variables are commonplace for example in digital image processing, where the picture elements (pixels) have always non-negative values. The cost functions and update rules of the MoG and rectified Gaussian node have been derived in Harva04DI. We postpone a more detailed discussion of these nodes to forthcoming papers to keep the length of this paper reasonable.
In the early conference paper Valpola01ICA where we introduced
the blocks for the first time, two more blocks were proposed for
handling discrete models and variables. One of them is a switch, which
picks up its 
-th continuous valued input signal as its output
signal. The other one is a discrete variable , which has a
soft-max prior derived from the continuous valued input signals 
of the node. However, we have omitted these two nodes from the present paper, because their
performance has not turned out to be adequate.
The reason might be that
assuming all parents of all nodes independent is too restrictive. For
instance, building a mixture-of-Gaussians from discrete and Gaussian
variables with switches is possible, but the construction loses out to
a specialised MoG node that makes fewer assumptions. In
Raiko05ICANN, the discrete node is used without switches.
Action and utility nodes Pearl88, Murphy01 would extend the library into decision theory and control. In addition to the messages about the variational Bayesian cost function, the network would propagate messages about utility. Raiko05IJCNN describe such a system in a slightly different framework.