Quadratic pseudo-Boolean optimization: Difference between revisions

Revision as of 21:22, 9 October 2020

Latest revision as of 10:19, 13 June 2024

(12 intermediate revisions by 7 users not shown)

Line 1:

{{short description|Combinatorial optimization method for pseudo-Boolean functions}}

'''Quadratic pseudo-Boolean optimisation''' ('''QPBO''') is a [[combinatorial optimization]] method for minimizing quadratic [[pseudo-Boolean function]]s in the form

:<math> f(\mathbf{x}) = w_0 + \sum_{p \in V} w_p(x_p) + \sum_{(p, q) \in E} w_{pq}(x_p, x_q) </math>

in the [[binary data|binary variables]] <math>x_p \in \{0, 1\} \; \forall p \in V = \{1, \dots, n\}</math>, with <math>E \subseteq V \times V</math>. If <math>f</math> is [[Submodular set functionPseudo-Boolean_function#Submodularity|submodular]] then QPBO produces a global optimum equivalently to [[graph cut optimization]], while if <math>f</math> contains non-submodular terms then the algorithm produces a partial solution with specific optimality properties, in both cases in [[polynomial time]].<ref name="review" />

QPBO is a useful tool for inference on [[Markov random field]]s and [[conditional random field]]s, and has applications in [[computer vision]] problems such as [[image segmentation]] and [[stereo cameras|stereo matching]].<ref name="rother" />

Line 31:

The algorithm can be divided in three steps: graph construction, max-flow computation, and assignment of values to the variables.

When constructing the graph, the set of vertices <math>V</math> contains the source and sink nodes <math>s</math> and <math>t</math>, and a couplepair of nodes <math>p</math> and <math>p'</math> for each variable. After re-parametrising the function to normal form,<ref name="normal form" group="note" /> a pair of edges is added to the graph for each term <math>w</math>:

* for each term <math>w_p(0)</math> the edges <math>p \rightarrow t</math> and <math>s \rightarrow p'</math>, with weight <math>\frac{1}{2} w_p(0)</math>;

* for each term <math>w_p(1)</math> the edges <math>s \rightarrow p</math> and <math>p' \rightarrow t</math>, with weight <math>\frac{1}{2} w_p(1)</math>;

Line 47:

== Higher order terms ==

The problem of optimizing higher-order pseudo-boolean functions is generally difficult. The process of reducing a high-order function to a quadratic one is known as "quadratization".<ref name="dattani" /> It is always possible to reduce a higher-order function to a quadratic function which is equivalent with respect to the optimisation, problem known as "higher-order [[clique (graph theory)|clique]] reduction" (HOCR), and the result of such reduction can be optimized with QPBO. Generic methods for reduction of arbitrary functions rely on specific substitution rules and in the general case they require the introduction of auxiliary variables.<ref name="fix" /> In practice most terms can be reduced without introducing additional variables, resulting in a simpler optimization problem, and the remaining terms can be reduced exactly, with addition of auxiliary variables, or approximately, without addition of any new variable.<ref name="elc" />

==Notes==

<references>

<ref name="dattani">Dattani (2019).</ref>

<ref name="review">Kolmogorov and Rother (2007).</ref>

<ref name="fix">Fix et al. (2011).</ref>

Line 61 ⟶ 60:

==References==

* {{cite journal|first1=Alain|last1=Billionnet|first2=Brigitte|last2=Jaumard|author2-link= Brigitte Jaumard |title=A decomposition method for minimizing quadratic pseudo-boolean functions|journal= [[Operations Research Letters]] |volume=8|number=3| pages=161–163|year=1989|doi=10.1016/0167-6377(89)90043-6}}

* {{cite arXiv|title=Quadratization in discrete optimization and quantum mechanics|eprint=1901.04405|first=Nike|last=Dattani|author-link=Nike Dattani|year=2019|class=quant-ph}}

* {{cite conference|first1=Alexander|last1=Fix|first2=Aritanan|last2=Gruber|first3=Endre|last3=Boros|first4=Ramin|last4=Zabih|title=A graph cut algorithm for higher-order Markov random fields|conference= [[International Conference on Computer Vision]] |year=2011|pages=1020–1027|url=https://www.cs.cornell.edu/~afix/Papers/ICCV11.pdf}}

* {{cite conference|first1=Hiroshi|last1=Ishikawa|title=Higher-Order Clique Reduction Without Auxiliary Variables|conference= [[Conference on Computer Vision and Pattern Recognition]]|year=2014|publisher=IEEE|pages=1362–1269|url=https://www.cv-foundation.org/openaccess/content_cvpr_2014/papers/Ishikawa_Higher-Order_Clique_Reduction_2014_CVPR_paper.pdf}}

Line 69 ⟶ 67:

== Notes ==

<references group="note">

<ref name="normal form">The representation of a pseudo-Boolean function with coefficients <math>\mathbf{w} = (w_0, w_1, \dots, w_{nn})</math> is not unique, and if two coefficient vectors <math>\mathbf{w}</math> and <math>\mathbf{w}'</math> represent the same function then <math>\mathbf{w}'</math> is said to be a reparametrisation of <math>\mathbf{w}</math> and vice versa. In some constructions it is useful to ensure that the function has a specific form, called ''noramlnormal form'', which is always defined for any function, and it is not unique. A function <math>f</math> is in normal form if the two following conditions hold (Kolmogorov and Rother (2007)):

# <math>\min \{ w_p^0, w_p^1 \} = 0</math> for each <math>p \in V</math>;

# <math>\min \{ w_{pq}^{0j}, w_{pq}^{1j} \} = 0</math> for each <math>(p, q) \in E</math> and for each <math>j \in \{0, 1\}</math>.