Parallel MCE reasoning and Boltzmann-Jeffrey machine networks

A parallel computational model reasoning under uncertainty, i.e. a network of Boltzmann-Jeffrey machines, is proposed based on an application of the principle of minimum cross entropy (MCE) to recursive causal models. The efficiency of the proposed model is analyzed, and a simple comparison is given for this model and some other similar models, such as Bayesian networks and connectionist networks. The proposed model can be used not only in conventional RCNets (recursive causal networks) but also in some more general networks. There may be even some small directed cycles included inside each RCNDL (recursive causal network description language) clause if only they do not violate the conditional independence among the clauses. In addition to Jeffrey´s rule, some more general minimum cross entropy techniques can also be included. Reasoning in all directions is allowed. The result obtained by this method can be guaranteed to be as accurate as that obtained with the Bayesian method. The method is quite efficient for applications with large sparse probabilistic spaces if sufficient hardware resources are available