The best understood approach for modeling continuous distributions
in Bayesian network learning is based on Gaussian
distributions [11]. This form of continuous Bayesian
network can be learned using exact Bayesian derivations
quite efficiently. Unfortunately, the expressive power of
Gaussian networks is limited. Formally, "pure" Gaussian
networks can only learn linear dependencies among the
measured variables.
This is a serious restriction when learning in domains with
non-linear interactions, or domains where the nature of the
interactions is unknown. A common way of avoiding this
problem is to introduce hidden variables that represent mixtures
of Gaussians (e.g., [28, 34]). An alternative approach
that has been suggested is to learn with non-parametric densities
[17].