How to Compute a Primal Solution From Dual One in LP Relaxation of MAP Inference in MPF?

Описан метод получения оптимальной размытой разметки по оптимальному решению двойственной задачи LP-релаксации на марковских случайных полях. Метод основан на LP-релаксации специального вида и алгоритме (max,+)-диффузии....

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Datum:2011
1. Verfasser: Werner, T.
Format: Artikel
Sprache:English
Veröffentlicht: Міжнародний науково-навчальний центр інформаційних технологій і систем НАН та МОН України 2011
Schriftenreihe:Управляющие системы и машины
Schlagworte:
Online Zugang:https://nasplib.isofts.kiev.ua/handle/123456789/82927
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Назва журналу:Digital Library of Periodicals of National Academy of Sciences of Ukraine
Zitieren:How to Compute a Primal Solution From Dual One in LP Relaxation of MAP Inference in MPF? / T. Werner // Управляющие системы и машины. — 2011. — № 2. — С. 86-93. — Бібліогр.: 20 назв. — англ.

Institution

Digital Library of Periodicals of National Academy of Sciences of Ukraine
id nasplib_isofts_kiev_ua-123456789-82927
record_format dspace
spelling nasplib_isofts_kiev_ua-123456789-829272025-02-23T17:52:38Z How to Compute a Primal Solution From Dual One in LP Relaxation of MAP Inference in MPF? Как находить решение прямой задачи LP-релаксации MRF по решению двойственной задачи ? Як обчислювати розв’язок прямої задачі LP-релаксації MRF за розв’язком двоїстої задачі ? Werner, T. Оптимизационные задачи структурного распознавания образов Описан метод получения оптимальной размытой разметки по оптимальному решению двойственной задачи LP-релаксации на марковских случайных полях. Метод основан на LP-релаксации специального вида и алгоритме (max,+)-диффузии. The method for computation of the optimal relaxed labeling from the optimal solution of a dual problem in the LP-relaxation of Markov Random Fields is described. The method is based on the particular form of the LP-relaxation and the (max,+)-diffusion algorithm. Описано метод отримання оптимальної розмитої розмітки за оптимальним розв’язком двоїстої задачі LP-релаксації на марківських випадкових полях. Метод базується на LP-релаксації спеціального типу і алгоритмі (max,+)-дифузії. This research has been supported by the European Community project FP7-ICT-247022 (MASH) and by the Czech government grant MSM6840770038. The author thanks Sebastian Nowozin for useful discussions. 2011 Article How to Compute a Primal Solution From Dual One in LP Relaxation of MAP Inference in MPF? / T. Werner // Управляющие системы и машины. — 2011. — № 2. — С. 86-93. — Бібліогр.: 20 назв. — англ. 0130-5395 https://nasplib.isofts.kiev.ua/handle/123456789/82927 519.157 en Управляющие системы и машины application/pdf Міжнародний науково-навчальний центр інформаційних технологій і систем НАН та МОН України
institution Digital Library of Periodicals of National Academy of Sciences of Ukraine
collection DSpace DC
language English
topic Оптимизационные задачи структурного распознавания образов
Оптимизационные задачи структурного распознавания образов
spellingShingle Оптимизационные задачи структурного распознавания образов
Оптимизационные задачи структурного распознавания образов
Werner, T.
How to Compute a Primal Solution From Dual One in LP Relaxation of MAP Inference in MPF?
Управляющие системы и машины
description Описан метод получения оптимальной размытой разметки по оптимальному решению двойственной задачи LP-релаксации на марковских случайных полях. Метод основан на LP-релаксации специального вида и алгоритме (max,+)-диффузии.
format Article
author Werner, T.
author_facet Werner, T.
author_sort Werner, T.
title How to Compute a Primal Solution From Dual One in LP Relaxation of MAP Inference in MPF?
title_short How to Compute a Primal Solution From Dual One in LP Relaxation of MAP Inference in MPF?
title_full How to Compute a Primal Solution From Dual One in LP Relaxation of MAP Inference in MPF?
title_fullStr How to Compute a Primal Solution From Dual One in LP Relaxation of MAP Inference in MPF?
title_full_unstemmed How to Compute a Primal Solution From Dual One in LP Relaxation of MAP Inference in MPF?
title_sort how to compute a primal solution from dual one in lp relaxation of map inference in mpf?
publisher Міжнародний науково-навчальний центр інформаційних технологій і систем НАН та МОН України
publishDate 2011
topic_facet Оптимизационные задачи структурного распознавания образов
url https://nasplib.isofts.kiev.ua/handle/123456789/82927
citation_txt How to Compute a Primal Solution From Dual One in LP Relaxation of MAP Inference in MPF? / T. Werner // Управляющие системы и машины. — 2011. — № 2. — С. 86-93. — Бібліогр.: 20 назв. — англ.
series Управляющие системы и машины
work_keys_str_mv AT wernert howtocomputeaprimalsolutionfromdualoneinlprelaxationofmapinferenceinmpf
AT wernert kaknahoditʹrešenieprâmojzadačilprelaksaciimrfporešeniûdvojstvennojzadači
AT wernert âkobčislûvatirozvâzokprâmoízadačílprelaksacíímrfzarozvâzkomdvoístoízadačí
first_indexed 2025-11-24T05:53:25Z
last_indexed 2025-11-24T05:53:25Z
_version_ 1849649901269417984
fulltext 86 УСиМ, 2011, № 2 UDC 519.157 T. Werner How to Compute a Primal Solution from Dual One in LP Relaxation of the MAP Inference in MRF ? Описан метод получения оптимальной размытой разметки по оптимальному решению двойственной задачи LP-релаксации на марковских случайных полях. Метод основан на LP-релаксации специального вида и алгоритме (max,+)-диффузии. The method for computation of the optimal relaxed labeling from the optimal solution of a dual problem in the LP-relaxation of Markov Ran- dom Fields is described. The method is based on the particular form of the LP-relaxation and the (max,+)-diffusion algorithm. Описано метод отримання оптимальної розмитої розмітки за оптимальним розв’язком двоїстої задачі LP-релаксації на марків- ських випадкових полях. Метод базується на LP-релаксації спеціального типу і алгоритмі (max,+)-дифузії. Abstract In the LP relaxation of MAP inference in Mar- kov random fields (MRF), the primal LP maximi- zes the MAP objective over relaxed labelings (pseu- domarginals) and the dual LP minimizes an upper bound on the true MAP solution by reparameteri- zations. Having solved the dual LP, we have no direct access to the corresponding primal solution. We propose a simple way to compute an optimal primal solution from an optimal dual solution. Precisely, we give an algorithm that either shows that the upper bound for a given problem can be further decreased by reparameterizations (i.e., it is not dual-optimal) or computes the corresponding optimal relaxed labeling. This is done by first re- moving inactive dual constraints and then solving the resulting feasibility problem by a very simple message-passing algorithm, sum-product diffusion. The MAP inference in undirected graphical mo- dels (Markov random fields, MRF) [16] leads to the following NP-hard combinatorial optimization problem: given a set of variables and a set of func- tions of (small) subsets of the variables, maximize the sum of the functions over all the variables. The problem has a natural LP relaxation, proposed independently in [15, 8, 16]. The primal task in this LP relaxation maximizes the MAP objective over relaxed labelings (pseudomarginals), the dual LP minimizes an upper bound on the true MAP solution by reparameterizations (equivalent trans- formations) of the original problem. Currently, the only known algorithms able to compute the LP relaxation for large-scale instan- ces are dual. Examples are algorithms based on ave- raging max-marginals [10, 6, 4, 5], the Augmen- ting DAG/VAC algorithm [9, 1], subgradient me- thods [7, 14], or smoothing approaches [20, 17, 5, 12]. Having an optimal dual solution, it is not easy to compute the corresponding primal solution (op- timal relaxed labeling) for large-scale instances. However, this primal solution can be sometimes useful. In this paper, we present a simple algo- rithm to compute an optimal primal solution from an optimal dual solution in the LP relaxation. Pre- cisely, suppose somebody gives us a problem and claims it has the minimal upper bound among all its reparameterizations. We want to either dis- prove this claim or compute the corresponding pri- mal optimum. Given a dual solution, we first remove inactive dual constraints, which means setting the corre- sponding primal variables (pseudomarginals) to zero. We are left with a feasibility task. This fea- sibility task is replaced with its a smooth optimi- zation task such that the primal optimum of the smoothed task is a solution of the feasibility task. The optimum of the smoothed task can be com- puted with a very simple message passing algo- rithm, which we call sum-product diffusion. If the dual solution was not optimal, this is detected dur- ing sum-product diffusion and a decreasing direc- tion for the upper bound is provided. We build our paper around the particular form of the LP relaxation proposed in [15] and the max- sum diffusion algorithm [10], which were reviewed in [18]. Since researchers are most familiar with problems with pairwise interactions, we present our algorithm on these – but it can be straightfor- wardly extended to the LP relaxation of problems of an arbitrary order (arity), as proposed in [19, 3]. УСиМ, 2011, № 2 87 1. The MAP inference in MRF and its LP re- laxation Let V be a set of variables and      2 VE a set of variable pairs, thus (V, E) is a graph. Variable u V attains states xu from a finite domain Xu. An as- signment (or labeling) is a tuple x X, where X is the Cartesian product of the domains Xu for all u V. We want to maximize the function )],()([=)|( vuuv Euv uu Vu xxfxffxF    (1) over all x X, where the functions : [ , )u uf X    and ),[:  vuuv XXf are given. All numbers )( uu xf and ),( vuuv xxf form a single vector Tf ),[  , where },,|),({ },|),({= vvuuvu uuu XxXxEuvxxuv XxVuxuT   denotes the set of all states of all the functions. We understand E as an undirected graph, adopting that ),(=),( uvvuvuuv xxfxxf . We will need also the directed version E of E, such that for any undi- rected edge uv E we have arcs in both directions uv E and vu E. Here are the primal (left) and the dual (right) of the LP relaxation of the problem [15, 18]: .)(min=,max    fUf (2) In the primal, T[0,1] is the set of vectors T[0,1] satisfying uuuuvuuv vx XxEuvxxx  ,,)(=),( * (3a) ( ) = 1 , . u u u x x u V  (3b) The set  is often referred to as the local mar- ginal polytope and the components of  as pseu- domarginals [16]. A vector  is the collection of distributions u and uv, subject to marginali- zation constraints (3a) and normalization con- straints (3b). In the scalar product  ,f we de- fine – ¥  0 = 0. The meaning of the dual can be understood by combining two concepts, reparameterizations and an upper bound on (1) , as shown next. A reparameterization is a linear transformation of vector f that preserves )|( fxF for all x X. The simplest reparameterization is done as follows: pick any uv E, subtract an arbitrary unary function (usually called a «message») ),(:  uuv X from function fu and add it to function fuv: ( ) ( ) ( )u u u u uv uf x f x x  (4a) .)(),(),( uuvvuuvvuuv xxxfxxf  (4b) This preserves uvu ff  and hence )|( fF  . Ap- plying to all *Euv yields )()(=)( uuv uNv uuuu xxfxf     (5a) ,)()(),(=),( vvuuuvvuuvvuuv xxxxfxxf  (5b) where }|{= EuvvNu  is the set of neighbors of variable u. Vector  is the collection of all the numbers )( uuv x , thus f  denotes the reparame- terization of vector f by messages  . Reparame- terizations preserve not only )|( fF  but also the primal objective, i.e., for any  we have  ,=, ff . The function U in the dual is an upper bound on )|( fF  , .)|(max ),(max)(max=)( , fxF xxfxffU Xx vuuv vxuxEuv uu uxVu      (6) The meaning of the dual is now evident: it minimizes the upper bound (6) by reparameteriza- tions (5). We will say that f is a dual optimum if it has the least upper bound among all its reparameteri- zations, i.e., U (f ) = min U (f  ). Note that this is a slight abuse of terminology because the decision variables in the dual in (2) are  rather than f. 1.1. The Relations between the primal and dual The primal and dual are related by the well- known duality theorems:  Weak duality: For any    and any f, we have )(, fUf  .  Strong duality: For   , we have , =f  = U (f ) if and only if  is a primal optimum and f is a dual optimum. 88 УСиМ, 2011, № 2  Complementary slackness: For   , we have )(=, fUf  if and only if [ ( ) ( )] ( ) = 0max u u u u u u yu f y f x x  (7a) , [ ( , ) ( , )] ( ) = 0 .max uv u v uv u v uv uv y yu v f y y f x x x  (7b) If f is not a dual optimum then there exists no    satisfying (7). 1.2. Computing a primal solution as a feasi- bility problem Our task in this paper is to compute a primal optimum from a dual optimum. More precisely, somebody gives us a vector f and claims it has the minimal upper bound among all its reparameteri- zations, U (f ) = min U (f  ). Our task is either to disprove this claim or to compute    such that )(=, fUf  . Let us replace the given vector f  [– , )T with vector g  {– , 0}T defined as      )(max=)(0 )(max<)( =)( uu uy uu uu uy uu uu yfxfif yfxfif xg , (8a) . ),(max =),( 0 ),(max <),( =),( , ,           vuuv vyuy vuuv vuuv vyuy vuuv vuuv yyf xxfif yyf xxfif xxg (8b) We assume that >)( fU , hence 0=)(gU . Replacing f with g can be seen as removing inac- tive dual constraints and setting the corresponding primal variables to zero. This does not change our task. Two cases can arise:  f is not a dual optimum: Clearly, f is a dual optimum if and only if g is a dual optimum. Hence, there exists  such that U (g ) < 0. This  is at the same time a decreasing direction for U (f ), i.e., there exists 0> such that )(<)( fUfU  ;  f is a dual optimum: Clearly, if f is a dual op- timum then problems f and g have the same set of optimal primal solutions. Hence, there exists  satisfying (7) , which is the desired primal optimum. Condition (7) can be written in short as 0=, g , recalling that 0=0 . Now, our task in this paper reduces to the fol- lowing feasibility problem: Problem 1 Given g  {– , 0}T, find    sa- tisfying 0=, g . If such  does not exist, find  such that 0<)( gU . 2. Solving the feasibility problem Here we describe a solution to Problem 1. In short, the idea is to replace task (2) with its smo- othed version such that for g  {– , 0}T, the pri- mal optimum of the smoothed task is a primal op- timum of the original task. The (dual and primal) global optimum of the smoothed task can be com- puted with a very simple message passing algo- rithm. Although the smoothed version of (2) will be eventually applied to g  {– , 0}T, we formulate it for g  [– , 0)T. It reads .)( ~ min=,logmax    gUg (9) The primal (left) can be understood as a mini- mization of a convex free energy (here, maximiza- tion of negative energy). In the dual (right), we have )|(),( )(=)( ~ , gxFxxg xggU Xx vuuv vxuxEuv uu uxVu       (10) where = log expi ii i a a denotes the log-sum-exp operation. The right-hand expression in (10) is the log-partition function, hence U ~ is an upper bound on the log-partition function. The bound is too loose to be useful for approximating (log-) parti- tion function but this is not our task here. The dual is a differentiable convex task, it is at optimum if )(=),( uuvuuv vx xgxxg  (11) holds for all *Euv and ux . The primal and dual optima are related by , exp ( ) ( ) = , exp ( ) exp ( , ) ( , ) = . exp ( , ) u u u u u u xu uv u v uv u v uv u v x xu v g x x g x g x x x x g x x         (12) УСиМ, 2011, № 2 89 We will not prove here the duality relation (9), the inequality in (10), and the optimality condition (11). This is easy and it can be found, e.g., in [20]. Plugging (12) into (11) verifies the marginaliza- tion condition (3a), hence   . We can see that to solve the dual task in (9), we need to reparameterize g to satisfy the stationary condition (11). Then the primal solution can be read off from (12). Note, the approach has the desirable property that pseudomarginals  need not be explicitly stored in the memory, they are given implicitly by (11). Therefore we need to store only unary func- tions  rather than unary and binary functions  . 2.1. The enforcing arc consistency It can happen that the condition (11) is impos- sible to satisfy by any choice of . This is because reparameterizations cannot change a finite weight to an infinite one or vice versa (note, messages cannot take infinite values) – in other words,  >)( uu xg if and only if >)( uu xg , and similarly for guv. Therefore, the finite part of g has to satisfy the property known as arc consistency [2]: for all uv E and xu we have ]>)(>),(max[  uuvuuv vx xgxxg . (13) Polynomial algorithms exist that recursively set some of the weights g to  until g becomes arc consistent. This is known as a relaxation labeling [13] or, in constraint satisfaction, as an enforcing arc consistency or a constraint propagation [2]. It is outlined in Algorithm 1. Algorithm 1. Enforcing arc consistency. 1: repeat; 2: Find uv E and xu violating (13). 3: )( uu xg 4: ),( vuuv xxg for all xv 5: until no change is possible It can be shown that enforcing arc consistency does not change the optimum of (9). In particular, if Algorithm 1 sets all weights to  ( =g ), Problem 1 is infeasible. 2.2. A message passing algorithm The dual in (9) can be solved by coordinate de- scent, which leads to a message passing algorithm. The algorithm repeats the following iteration until convergence: Pick any uv E and enforce equality (11) by reparameterization (4). This strictly monotonically decreases )( ~ gU . On convergence, (11) holds globally. Algorithm 2. The ),(  -diffusion algorithm. 1: repeat; 2: for *Euv and uu Xx  >)( uu xg do 3:             ),()( 2 1 )()( vuuv vx uu uuvuuv xxgxg xx 4: end for 5: until convergence This is summarized in the Algorithm 2. It is pre- cisely analogical to max-sum diffusion [10, 18], only the function max was replaced with the log- sum-exp function . The algorithm assumes that g is arc consistent – then, the condition >)( uu xg ensures that )( uuv x is never set to  or  . 2.3. Solving the feasibility problem Let us put things together and see how to solve Problem 1. For Tg ,0}{ , any  satisfying  >,logg satisfies also 0=, g . Hence, any  feasible to the primal in (9) solves Problem 1. For any g we have )()( ~ gUgU  because the log-sum-exp function upper bounds the maxi- mum,  i ai  maxi ai, Hence, if during the Algo- rithm 2 we get 0<)( ~ gU , it implies 0<)( gU . Thus, Problem 1 is solved as follows: 1. Enforce arc consistency, Algorithm 1. If =g , stop (Problem 1 is infeasible); 2. Run Algorithm 2. If 0<)( ~ gU any time dur- ing the algorithm, stop (Problem 1 is infeasible); 3. Otherwise, let Algorithm 2 converge and then compute a solution  of Problem 1 from (12) . Before the convergence of Algorithm 2,  given by (17) satisfies 0=, g but violates the mar- ginalization constraint (3a). On convergence,  satisfies (3b), hence  . 90 УСиМ, 2011, № 2 a b Fig. 1. The depicted reparameterization decreases edge e to an arbitrarily small value  2.4. Unbounded messages The algorithm has an interesting but undesired property: even if problem the (9) is bounded, its optimum may be attained for  . In that case, during the Algorithm 2 some of the messa- ges )( uuv x will diverge to  or  . Fig. 1, a shows an example of this phenome- non. It shows a problem with 3=V variables, the complete graph E, and 2|=| uX labels. All the nodes have weights 0=)( uu xg . The shown edges have weights 0=),( vuuv xxg , the remaining (not shown) edges have weights =),( vuuv xxg . Consider the reparameterization  given by setting  =)( uuv x for the six messages depicted by dashes, and 0=)( uuv x for the remaining six messages. For any  , this reparameterization changes the weight of edge e to  =)(eg and leaves the remaining weights unchanged. For any 0 , we have =)( gU ( ) = 0U g . By comple- mentary slackness, it necessarily follows that 0=)(e . Although the bound )( gU is the same for any 0 , the smoothed upper bound )( ~ gU de- creases with increasing . Since Algorithm 2 minimizes )( ~ gU , it will keep increasing  (and hence decreasing )(eg ) without any limits. This must be so because by (12), for 0=)(e we need  =)(eg . This behavior can lead to numerical problems because in (5) we can get expressions like  ),(=),( vuuvvuuv xxgxxg where  is a very large number. This can be alleviated by doing re- parameterizations (1) «in-place» by directly chang- ing g (Algorithm 3) rather than by storing mes- sages – but in that case we may instead have prob- lems with error accumulation. Moreover, modify- ing binary functions ),( uvg typically needs more memory than storing only unary functions )(uv . Algorithm 3. The ),(  -diffusion algorithm, message-free version. 1: repeat; 2: for *Euv and uu Xx  such that >)( uu xg do 3: )],()([ 2 1 vuuv vxuu xxgxg  4:  )()( uuuu xgxg 5:  ),(),( vuuvvuuv xxgxxg for all vx 6: end for 7: until convergence The phenomenon can be further clarified as fol- lows. Let two vectors Tgg ),[,  be called equivalent if they define the same function )|( gF  . An equivalent transformation is a change of vec- tor g to its equivalent. Now, three classes of equiva- lent transformations can be distinguished [20]: 1. Transformations that are compositions of a finite number of local reparameterizations (4). These are precisely the linear transformations (5). 2. Transformations that are compositions of an infinite number of local reparameterizations (4). The resulting transformations are not all covered by (5) because (5) does not allow to change a fi- nite weight to  or vice versa. 3. Transformations that are not compositions of any number of local reparameterizations (4). E.g., any unsatisfiable CSP is equivalent to the empty CSP. The problems in Figure 1 (a) and 1 (b) are equi- valent in the second sense. Hence, there exists no  that would reparameterize Figure 1 (a) to Fig- ure 1 (b) or vice versa. 2.5. A sum-product version of the algorithm The Algorithms 2 and 3 require a time-consu- ming evaluation of the log-sum-exp function. By applying the exponential function to all involved УСиМ, 2011, № 2 91 quantities, these algorithms (and the theory in § 2) can be translated from the semiring ),),,([  into the sum-product semiring ),),([0,  . Then, the algorithms will use only addition, multiplica- tion, division, and square root. We refer to the re- sulting algorithm as sum-product diffusion. Its message-free version is Algorithm 4. Its input is a vector g  [0, )T that is assumed to be arc con- sistent, i.e., 0>),(max vuuvvx xxg if and only if 0>)( uu xg . Algorithm 4. The sum-product diffusion algo- rithm, message-free version. 1: repeat; 2: for *Euv and uu Xx  such that >)( uu xg do 3: 1/2)],()/([ vuuv vxuu xxgxg  4:  )/()( uuuu xgxg 5:  ),(),( vuuvvuuv xxgxxg for all vx 6: end for 7: until convergence In the sum-product form, the core idea of our paper is especially obvious. Inequality (10) reads ,)|( ),()(=)( ~ , gxF xxgxggU Xx vuuv vxuxEuv uu uxVu                       (14) where .),()(=)|( vuuv Euv uu Vu xxgxggxF   (15) The stationary condition (11) reads .)(=),( uuvuuv vx xgxxg (16) Now, the problem 1 is solved simply by apply- ing sum-product diffusion to g  {0, 1}T. It is ob- vious that after convergence, the optimal  coin- cides with g (up to normalization) – because (16) is the marginalization condition (3a) we want to impose. Moreover, the test for infeasibility of Prob- lem 1 has an interesting interpretation in terms of the constraint satisfaction problem (CSP) [11]. A vector g  {0, 1}T can be understood as a (pair- wise) CSP, the functions gu : Xu  {0, 1} and {0,1}:  vuuv XXg representing unary and bi- nary relations. A solution of the CSP is an as- signment Xx satisfying all the relations, i.e., 1=)|( gxF . The CSP is satisfiable if it has at least one solution. For g  {0, 1}T, the right-hand side of inequal- ity (14) is the number of solutions of the CSP. This is known as the counting constraint satisfac- tion problem (#CSP). Thus, )( ~ gU is an upper bound on the number of solutions of the CSP. This bound is too loose to be useful, except in one situation: 1<)( ~ gU implies that the CSP is unsat- isfiable. Algorithm 4 minimizes )( ~ gU by re- parameterizing g, hence preserving )|( gF  . If any time during the algorithm we get 1<)( ~ gU , we know that the CSP is unsatisfiable. Therefore, sum-product diffusion provides a test to disprove satisfiability of a CSP. This test is dissimilar to all tests based on local consistencies [2], used in con- straint programming. However, note that it is not apparent at the first sight that this test is equiva- lent to satisfiability of the Problem 1 – to show this, we needed the duality relation (9) . 3. What about non-optimal dual solutions? Many algorithms to tackle the dual in (2) yield only suboptimal dual solutions f – most impor- tantly, the algorithms based on the averaging max- marginals, such as max-sum diffusion [10, 18] or TRW-S [6]. An example of a problem that is a fixed point of these algorithms but is not a dual optimum is in Fig. 2. In that case, Algorithm 2 does not converge and achieves 0<)( ~ gU . Therefore, it would be useful to compute a subop- timal primal solution from a suboptimal dual solu- tion. We do not know how to modify our algo- rithm in a principled way to achieve this. But one can think of several heuristics. Of the duality relations (§ 1.1), one can sacri- fice either primal feasibility or zero duality gap. One option is to minimize primal infeasibility (viola- tion of (3a) ) subject to zero duality gap – but we do not know how to do this for large instances. 92 УСиМ, 2011, № 2 Fig. 2. An example of a fixed point of max-sum diffusion (or TRW-S) that is not a dual optimum (given by Schlesinger [18]) However, our algorithm can be modified to find a feasible primal solution with possibly small (but in general not minimal) duality gap. Instead of (8), let g be defined by       )(max)(0 )(max<)( =)(2 uu uy uu uu uy uu uu yfxfif yfxfif xg (17a)            ),(max ),( 0 ),(max <),( =),( , , vuuv vyuy vuuv vuuv vyuy vuuv vuuv yyf xxfif yyf xxfif xxg . (17b) For any   0 and  > 0, the Algorithm 2 will converge (with either sign of )( ~ gU ) and (12) will be a feasible primal solution  . With a small  and large  , one can expect a reasonably small duality gap  ,)( ffU . Unfortunately, this cannot be guaranteed based on solely  and  because for unbounded weights g the gap can be un- bounded in general. Currently, we cannot provide a bound on the duality gap based on  ,  and g. 4. Conclusion We have proposed an algorithm to compute a primal optimal solution from an optimal dual solu- tion in the LP relaxation of MAP-MRF inference. If the dual solution is not optimal, we provide a decreasing direction as a certificate of non- optimality. The main idea of our approach is ex- tremely simple, in fact being summarized by § 2.5. We have discussed an interesting but unde- sirable behavior of the algorithm, unbounded mes- sages. In this paper, we do not provide enough empi- rical evidence that our approach is useful in prac- tice. We tested the algorithm on instances with a sparse (grid) graph E and random weights f drawn i.i.d. from the normal distribution. Dual solutions were obtained by the max-sum diffusion. For the in- stances with V = 202 variables, the dual solution was optimal in approx. 50% cases, for the instances with V = 502 variables almost never. For optimal dual so- lutions, optimal primal solutions could always be found up to a small primal infeasibility. Unboun- ded messages never caused a serious problem. We formulated the algorithm for one particular form of the LP dual [15, 18], closely related to max- sum diffusion [10, 18]. However, it can be easily extended to algorithms with tree-based updates, both using the max-marginal averaging [6] and subgra- dients [7, 14]. In that case, the problem g would have to be defined in a different way than by (8). Acknowledgment This research has been supported by the Euro- pean Community project FP7-ICT-247022 (MASH) and by the Czech government grant MSM6840770038. The author thanks Sebastian Nowozin for useful discussions. 1. Soft arc consistency revisited / M. Cooper, S. de Givry, M. Sanchez, T. Schiex, M. Zytnicki, T. Werner // Arti- ficial Intelligence. – 2010. – 174(7–8). – P. 449–478. 2. Debruyne R., Bessiиre Ch. Domain Filtering Consis- tencies // J. of Artificial Intelligence Research. – 2001. – 14. – P. 205–230. 3. Franc V., Sonnenburg S., Werner T. Cutting Plane Methods in Machine Learning / S. Sra, S. Nowozin, S.J. Wright, ed. // Optimization for Machine Learning. MIT Press, 2011. 4. Globerson A., Jaakkola T. Fixing Max-Product: Con- vergent Message Passing Algorithms for MAP LP- Relaxations // Neural Inform. Proc. Syst. (NIPS). – 2008. – P. 553–560. 5. Johnson J. Malioutov D.M., Willsky A.S. Lagrangian rela- xation for MAP estimation in graphical models // Al- lerton Conf. Communication, Control and Computing, 2007. 6. Kolmogorov V. Convergent Tree-Reweighted Message Passing for Energy Minimization // IEEE Trans. Pat- tern Analysis and Machine Intelligence. – 2006. – 28(10). – P. 1568–1583. 7. Komodakis N., Paragios N., Tziritas G. MRF Optimi- zation via Dual Decomposition: Message-Passing Re- visited // Intl. Conf. Computer Vision (ICCV), 2007. УСиМ, 2011, № 2 93 8. Koster A., van Hoesel C.P.M., Kolen A.W.J. The Par- tial Constraint Satisfaction Problem: Facets and Lifting Theorems // Operations Research Letters. – 1998. – 23(3–5). – P. 89–97. 9. Koval V.K., Schlesinger M.I. Two-dimensional Program- ming in Image Analysis Problems // Automatics and Telemechanics. – 1976. – 8. – P. 149–168. In Russian. 10. Kovalevsky V.A., Koval V.K. A diffusion algorithm for decreasing the energy of the max-sum labeling prob- lem. Glushkov Institute of Cybernetics, Kiev, USSR. Unpublished, approx. 1975. 11. Mackworth A. Constraint Satisfaction // Encyclopaedia of Artificial Intelligence, Wiley, 1991. – P. 285–292. 12. Ravikumar P., Agarwal A., Wainwright M.J. Message- passing for Graph-structured Linear Programs: Proxi- mal Methods and Rounding Schemes // J. of Machine Learning Research. – 2010. – 11. – P. 1043–1080. 13. Rosenfeld A., Hummel R.A., Zucker S.W. Scene Label- ing by Relaxation Operations// IEEE Trans. on Syst., Man, and Cybernetics. 1976. – 6(6). – P. 420–433. 14. Schlesinger M.I., Giginjak V.V. Solving (max,+) Prob- lems of Structural Pattern Recognition Using Equivalent Transformations // USiM (Control Systems and Ma- chines). – 2007. – N 1,2. In Russian, English transla- tion available on www.irtc.org.ua/image/publications. 15. Schlesinger M.I. Syntactic analysis of two-dimensional visual signals in noisy conditions // Cybernetics and Systems Analysis. – 1976. – 12(4). – P, 612–628. Translation from Russian. 16. Wainwright J., Jordan M.I. Graphical Models, Expo- nential Families, and Variational Inference // Foundations and Trends in Machine Learning. – 2008. – 1(1–2). – P. 1–305. 17. Weiss Y., Yanover Ch., Meltzer T. MAP Estimation, Line- ar Programming and Belief Propagation with Convex Free Energies // Conf. Uncertainty in Artificial Intelli- gence (UAI), 2007. 18. Werner T. A Linear Programming Approach to Max- sum Problem: A Review // IEEE Trans. Pattern Analy- sis and Machine Intelligence. – July 2007. – 29(7). – P. 1165–1179. 19. Werner T. Revisiting the Linear Programming Relaxa- tion Approach to Gibbs Energy Minimization and Wei- ghted Constraint Satisfaction // Ibid. – August 2010. 20. Werner T., Shekhovtsov A. Unified Framework for Semi- ring-Based Arc Consistency and Relaxation Labeling // 12th Computer Vision Winter Workshop, 2007. – P. 27–34. – St. Lambrecht, Austria, Graz University of Technology. © T. Werner, A Center for MachinePerception, Czech Technical University, Prague, 2011  << /ASCII85EncodePages false /AllowTransparency false /AutoPositionEPSFiles true /AutoRotatePages /None /Binding /Left /CalGrayProfile (Dot Gain 20%) /CalRGBProfile (sRGB IEC61966-2.1) /CalCMYKProfile (U.S. Web Coated \050SWOP\051 v2) /sRGBProfile (sRGB IEC61966-2.1) /CannotEmbedFontPolicy /Error /CompatibilityLevel 1.4 /CompressObjects /Tags /CompressPages true /ConvertImagesToIndexed true /PassThroughJPEGImages true /CreateJobTicket false /DefaultRenderingIntent /Default /DetectBlends true /DetectCurves 0.0000 /ColorConversionStrategy /CMYK /DoThumbnails false /EmbedAllFonts true /EmbedOpenType false /ParseICCProfilesInComments true /EmbedJobOptions true /DSCReportingLevel 0 /EmitDSCWarnings false /EndPage -1 /ImageMemory 1048576 /LockDistillerParams false /MaxSubsetPct 100 /Optimize true /OPM 1 /ParseDSCComments true /ParseDSCCommentsForDocInfo true /PreserveCopyPage true /PreserveDICMYKValues true /PreserveEPSInfo true /PreserveFlatness true /PreserveHalftoneInfo false /PreserveOPIComments true /PreserveOverprintSettings true /StartPage 1 /SubsetFonts true /TransferFunctionInfo /Apply /UCRandBGInfo /Preserve /UsePrologue false /ColorSettingsFile () /AlwaysEmbed [ true ] /NeverEmbed [ true ] /AntiAliasColorImages false /CropColorImages true /ColorImageMinResolution 300 /ColorImageMinResolutionPolicy /OK /DownsampleColorImages true /ColorImageDownsampleType /Bicubic /ColorImageResolution 300 /ColorImageDepth -1 /ColorImageMinDownsampleDepth 1 /ColorImageDownsampleThreshold 1.50000 /EncodeColorImages true /ColorImageFilter /DCTEncode /AutoFilterColorImages true /ColorImageAutoFilterStrategy /JPEG /ColorACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /ColorImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000ColorACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000ColorImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasGrayImages false /CropGrayImages true /GrayImageMinResolution 300 /GrayImageMinResolutionPolicy /OK /DownsampleGrayImages true /GrayImageDownsampleType /Bicubic /GrayImageResolution 300 /GrayImageDepth -1 /GrayImageMinDownsampleDepth 2 /GrayImageDownsampleThreshold 1.50000 /EncodeGrayImages true /GrayImageFilter /DCTEncode /AutoFilterGrayImages true /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /GrayImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000GrayACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000GrayImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasMonoImages false /CropMonoImages true /MonoImageMinResolution 1200 /MonoImageMinResolutionPolicy /OK /DownsampleMonoImages true /MonoImageDownsampleType /Bicubic /MonoImageResolution 1200 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000 /EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode /MonoImageDict << /K -1 >> /AllowPSXObjects false /CheckCompliance [ /None ] /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXOutputIntentProfile () /PDFXOutputConditionIdentifier () /PDFXOutputCondition () /PDFXRegistryName () /PDFXTrapped /False /CreateJDFFile false /Description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> /CHS <FEFF4f7f75288fd94e9b8bbe5b9a521b5efa7684002000410064006f006200650020005000440046002065876863900275284e8e9ad88d2891cf76845370524d53705237300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c676562535f00521b5efa768400200050004400460020658768633002> /CHT <FEFF4f7f752890194e9b8a2d7f6e5efa7acb7684002000410064006f006200650020005000440046002065874ef69069752865bc9ad854c18cea76845370524d5370523786557406300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c4f86958b555f5df25efa7acb76840020005000440046002065874ef63002> /CZE <FEFF005400610074006f0020006e006100730074006100760065006e00ed00200070006f0075017e0069006a007400650020006b0020007600790074007600e101590065006e00ed00200064006f006b0075006d0065006e0074016f002000410064006f006200650020005000440046002c0020006b00740065007200e90020007300650020006e0065006a006c00e90070006500200068006f006400ed002000700072006f0020006b00760061006c00690074006e00ed0020007400690073006b00200061002000700072006500700072006500730073002e002000200056007900740076006f01590065006e00e900200064006f006b0075006d0065006e007400790020005000440046002000620075006400650020006d006f017e006e00e90020006f007400650076015900ed007400200076002000700072006f006700720061006d0065006300680020004100630072006f00620061007400200061002000410064006f00620065002000520065006100640065007200200035002e0030002000610020006e006f0076011b006a016100ed00630068002e> /DAN <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> /DEU <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> /ESP <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> /ETI <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> /FRA <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> /GRE <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a stvaranje Adobe PDF dokumenata najpogodnijih za visokokvalitetni ispis prije tiskanja koristite ove postavke. Stvoreni PDF dokumenti mogu se otvoriti Acrobat i Adobe Reader 5.0 i kasnijim verzijama.) /HUN <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> /ITA <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> /JPN <FEFF9ad854c18cea306a30d730ea30d730ec30b951fa529b7528002000410064006f0062006500200050004400460020658766f8306e4f5c6210306b4f7f75283057307e305930023053306e8a2d5b9a30674f5c62103055308c305f0020005000440046002030d530a130a430eb306f3001004100630072006f0062006100740020304a30883073002000410064006f00620065002000520065006100640065007200200035002e003000204ee5964d3067958b304f30533068304c3067304d307e305930023053306e8a2d5b9a306b306f30d530a930f330c8306e57cb30818fbc307f304c5fc59808306730593002> /KOR <FEFFc7740020c124c815c7440020c0acc6a9d558c5ec0020ace0d488c9c80020c2dcd5d80020c778c1c4c5d00020ac00c7a50020c801d569d55c002000410064006f0062006500200050004400460020bb38c11cb97c0020c791c131d569b2c8b2e4002e0020c774b807ac8c0020c791c131b41c00200050004400460020bb38c11cb2940020004100630072006f0062006100740020bc0f002000410064006f00620065002000520065006100640065007200200035002e00300020c774c0c1c5d0c11c0020c5f40020c2180020c788c2b5b2c8b2e4002e> /LTH <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> /LVI <FEFF0049007a006d0061006e0074006f006a00690065007400200161006f00730020006900650073007400610074012b006a0075006d00750073002c0020006c0061006900200076006500690064006f00740075002000410064006f00620065002000500044004600200064006f006b0075006d0065006e007400750073002c0020006b006100730020006900720020012b00700061016100690020007000690065006d01130072006f00740069002000610075006700730074006100730020006b00760061006c0069007401010074006500730020007000690072006d007300690065007300700069006501610061006e006100730020006400720075006b00610069002e00200049007a0076006500690064006f006a006900650074002000500044004600200064006f006b0075006d0065006e007400750073002c0020006b006f002000760061007200200061007400760113007200740020006100720020004100630072006f00620061007400200075006e002000410064006f00620065002000520065006100640065007200200035002e0030002c0020006b0101002000610072012b00200074006f0020006a00610075006e0101006b0101006d002000760065007200730069006a0101006d002e> /NLD (Gebruik deze instellingen om Adobe PDF-documenten te maken die zijn geoptimaliseerd voor prepress-afdrukken van hoge kwaliteit. De gemaakte PDF-documenten kunnen worden geopend met Acrobat en Adobe Reader 5.0 en hoger.) /NOR <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> /POL <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> /PTB <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> /RUM <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> /RUS <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> /SKY <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> /SLV <FEFF005400650020006e006100730074006100760069007400760065002000750070006f0072006100620069007400650020007a00610020007500730074007600610072006a0061006e006a006500200064006f006b0075006d0065006e0074006f0076002000410064006f006200650020005000440046002c0020006b006900200073006f0020006e0061006a007000720069006d00650072006e0065006a016100690020007a00610020006b0061006b006f0076006f00730074006e006f0020007400690073006b0061006e006a00650020007300200070007200690070007200610076006f0020006e00610020007400690073006b002e00200020005500730074007600610072006a0065006e006500200064006f006b0075006d0065006e0074006500200050004400460020006a00650020006d006f0067006f010d00650020006f0064007000720065007400690020007a0020004100630072006f00620061007400200069006e002000410064006f00620065002000520065006100640065007200200035002e003000200069006e0020006e006f00760065006a01610069006d002e> /SUO <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> /SVE <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> /TUR <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> /UKR <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> /ENU (Use these settings to create Adobe PDF documents best suited for high-quality prepress printing. Created PDF documents can be opened with Acrobat and Adobe Reader 5.0 and later.) >> /Namespace [ (Adobe) (Common) (1.0) ] /OtherNamespaces [ << /AsReaderSpreads false /CropImagesToFrames true /ErrorControl /WarnAndContinue /FlattenerIgnoreSpreadOverrides false /IncludeGuidesGrids false /IncludeNonPrinting false /IncludeSlug false /Namespace [ (Adobe) (InDesign) (4.0) ] /OmitPlacedBitmaps false /OmitPlacedEPS false /OmitPlacedPDF false /SimulateOverprint /Legacy >> << /AddBleedMarks false /AddColorBars false /AddCropMarks false /AddPageInfo false /AddRegMarks false /ConvertColors /ConvertToCMYK /DestinationProfileName () /DestinationProfileSelector /DocumentCMYK /Downsample16BitImages true /FlattenerPreset << /PresetSelector /MediumResolution >> /FormElements false /GenerateStructure false /IncludeBookmarks false /IncludeHyperlinks false /IncludeInteractive false /IncludeLayers false /IncludeProfiles false /MultimediaHandling /UseObjectSettings /Namespace [ (Adobe) (CreativeSuite) (2.0) ] /PDFXOutputIntentProfileSelector /DocumentCMYK /PreserveEditing true /UntaggedCMYKHandling /LeaveUntagged /UntaggedRGBHandling /UseDocumentProfile /UseDocumentBleed false >> ] >> setdistillerparams << /HWResolution [2400 2400] /PageSize [612.000 792.000] >> setpagedevice