On the optimality of solutions of the max-product belief-propagation algorithm in arbitrary graphs

被引:314
作者
Weiss, Y [1 ]
Freeman, WT
机构
[1] Univ Calif Berkeley, Div Comp Sci, Berkeley, CA 94720 USA
[2] MERL, Cambridge, MA 02139 USA
基金
美国国家科学基金会;
关键词
Bayesian networks; belief propagation; maximum a posteriori (MAP) estimate; Markov random fields (MRFs); max-product; min-sum;
D O I
10.1109/18.910585
中图分类号
TP [自动化技术、计算机技术];
学科分类号
0812 ;
摘要
Graphical models, such as Bayesian networks and Markov random fields (MRFs), represent statistical dependencies of variables by a graph, The max-product "belief propagation" algorithm is a local-message-passing algorithm on this graph that is known to converge to a unique fixed point when the graph is a tree. Furthermore, when the graph is a tree, the assignment based on the fixed point yields the most probable values of the unobserved variables given the observed ones. Recently, good empirical performance has been obtained by running the max-product algorithm (or the equivalent min-sum algorithm) on graphs with loops, for applications including the decoding of "turbo" codes. Except for two simple graphs (cycle codes and single-loop graphs) there has been little theoretical understanding of the max-product algorithm on graphs with loops. Here we prove a result on the fixed points of max-product on a graph with arbitrary topology and with arbitrary probability distributions (discrete- or continuous-valued nodes). We show that the assignment based on a fixed point is a "neighborhood maximum" of the posterior probability: the posterior probability of the max-product assignment is guaranteed to be greater than all other assignments in a particular large region around that assignment. The region includes all assignments that differ from the max-product assignment in any subset of nodes that form no more than a single loop in the graph. In some graphs, this neighborhood is exponentially large. We illustrate the analysis with examples.
引用
收藏
页码:736 / 744
页数:9
相关论文
共 23 条
[1]   Iterative decoding on graphs with a single cycle [J].
Aji, SM ;
Horn, GB ;
McEliece, RJ .
1998 IEEE INTERNATIONAL SYMPOSIUM ON INFORMATION THEORY - PROCEEDINGS, 1998, :276-276
[2]   The generalized distributive law [J].
Aji, SM ;
McEliece, RJ .
IEEE TRANSACTIONS ON INFORMATION THEORY, 2000, 46 (02) :325-343
[3]  
Benedetto S., 1996, 42124 JPL TDA
[4]  
Bertsekas D. P., 1987, DYNAMIC PROGRAMMING
[5]  
FORNEY GD, 1998, 1998 INF THEOR WORKS
[6]   Learning low-level vision [J].
Freeman, WT ;
Pasztor, EC ;
Carmichael, OT .
INTERNATIONAL JOURNAL OF COMPUTER VISION, 2000, 40 (01) :25-47
[7]   Skewness and pseudocodewords in iterative decoding [J].
Frey, BJ ;
Kotter, R ;
Vardy, A .
1998 IEEE INTERNATIONAL SYMPOSIUM ON INFORMATION THEORY - PROCEEDINGS, 1998, :148-148
[8]   Signal-space characterization of iterative decoding [J].
Frey, BJ ;
Koetter, R ;
Vardy, A .
IEEE TRANSACTIONS ON INFORMATION THEORY, 2001, 47 (02) :766-781
[9]  
Gallager RG, 1963, LOW DENSITY PARITY C
[10]   STOCHASTIC RELAXATION, GIBBS DISTRIBUTIONS, AND THE BAYESIAN RESTORATION OF IMAGES [J].
GEMAN, S ;
GEMAN, D .
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 1984, 6 (06) :721-741