An efficient approach for finding the MPE in belief networks
This addresses a computational bottleneck in probabilistic reasoning for AI applications, offering a more general solution compared to prior methods limited to restricted topologies.
The paper tackles the problem of finding the I most probable explanations (MPE) in arbitrary belief networks, presenting a new approach that includes an algorithm for the first MPE and a linear-time algorithm for subsequent MPEs, and extends to subsets of variables.
Given a belief network with evidence, the task of finding the I most probable explanations (MPE) in the belief network is that of identifying and ordering the I most probable instantiations of the non-evidence nodes of the belief network. Although many approaches have been proposed for solving this problem, most work only for restricted topologies (i.e., singly connected belief networks). In this paper, we will present a new approach for finding I MPEs in an arbitrary belief network. First, we will present an algorithm for finding the MPE in a belief network. Then, we will present a linear time algorithm for finding the next MPE after finding the first MPE. And finally, we will discuss the problem of finding the MPE for a subset of variables of a belief network, and show that the problem can be efficiently solved by this approach.