Parallel processing algorithm for Bayesian network inference

Abstract

We introduce an algorithm for Bayesian network inference using parallel computations that perform variable-elimination over multiple threads of execution. The algorithm can be implemented on a collection of parallel execution entities on a single FPGA. Each execution entity performs addition and multiplication. Relative to the standard bucket elimination, the parallel algorithm reduces the computational time by an amount that depends on the coupling (probabilistic dependency) of the network and on the evidence available at time of prediction query.