In Ensemble Learning, the search for good models is sensitive to high probability mass and so the problems of over-fitting inherent to maximum likelihood and maximum posterior probability methods are removed.
The approximation of the posterior distribution assumes some degree of factorisation of the true distribution in order to make the approximation more tractable. Additionally the fixed form approximation also assumes some functional form of the factors.
It is often possible to affect the correctness of the approximation by the choice of parameterisation of the model. Also, the learning process tries to make the approximation more correct.
The free form approximation of the separable distribution will often result in Gaussian, Gamma, etc distributions if the parametrization of the model is chosen suitably. Therefore the optimisation process will suggest a parametrization for the problem.