Probabilistic inference is an attractive approach to uncertain reasoning and empirical learning in articial intelligence. Computational diculties arise, however, because probabilistic models with the necessary realism and exibility lead to complex distributions over high-dimensional spaces.