Programming the harmonium

The authors show how to synthesize, given a Bayesian network description of a probability distribution, an instance of P. Smolensky´s harmony network (1986) that computes maximum-likelihood completions to partial value assignments, according to the given distribution. As an application, they present a scheme for translating a high-level description of a conceptual hierarchy, with default values and exceptions, to a harmony network representation, which is then inherently capable of quite complicated query processing. Finally, the authors discuss their experience with implementing such a translation scheme