Smoothed Analysis of Belief Propagation for Minimum-Cost Flow and Matching
Belief propagation (BP) is a message-passing heuristic for statistical inference in graphical models such as Bayesian networks and Markov random fields. BP is used to compute marginal distributions or maximum likelihood assignments and has applications in many areas, including machine learning, image processing, and computer vision. However, the theoretical understanding of the performance of BP is unsatisfactory. Recently, BP has been applied to combinatorial optimization problems. It has been proved that BP can be used to compute maximum-weight matchings and minimum-cost flows for instances with a unique optimum. The number of iterations needed for this is pseudo-polynomial and hence BP is not efficient in general.
We study belief propagation in the framework of smoothed analysis and prove that with high probability the number of iterations needed to compute maximum-weight matchings and minimum-cost flows is bounded by a polynomial if the weights/costs of the edges are randomly perturbed. To prove our upper bounds, we use an isolation lemma by Beier and Vöcking (SIAM J. Comput., 2006) for matching and generalize an isolation lemma for min-cost flow by Gamarnik, Shah, and Wei (Oper. Res., 2012). We also prove almost matching lower tail bounds for the number of iterations that BP needs to converge.
KeywordsBipartite Graph Belief Propagation Complete Bipartite Graph Arbitrary Graph Graph Cover
Unable to display preview. Download preview PDF.
- 1.Ahuja, R.K., Magnanti, T.L., Orlin, J.B.: Network flows: theory, algorithms, and applications. Prentice-Hall (1993)Google Scholar
- 8.Pearl, J.: Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann (1988)Google Scholar
- 14.Yedidia, J.S., Freeman, W.T., Weiss, Y.: Understanding belief propagation and its generalizations. In: Lakemeyer, G., Nebel, B. (eds.) Exploring Artificial Intelligence in the New Millennium, ch. 8, pp. 239–269. Morgan Kaufmann (2003)Google Scholar