Investigating decomposition methods for the maximum common subgraph and sum colouring problems - Archive ouverte HAL Access content directly
Theses Year : 2017

Investigating decomposition methods for the maximum common subgraph and sum colouring problems

Utilisation de méthodes de décomposition pour les problèmes du plus grand sous-graphe commun et de la somme coloration


The objective of this thesis is, from a general standpoint, to design and evaluate decomposition methods for solving constrained optimisation problems. Two optimisation problems in particular are considered: the maximum common induced subgraph problem, in which the largest common part between two graphs is to be found, and the sum colouring problem, where a graph must be coloured in a way that minimises a sum of weights induced by the employed colours. The maximum common subgraph (MCIS) problem is notably difficult, with a strong applicability in domains such as biology, chemistry and image processing, where the need to measure the similarity between structured objects represented by graphs may arise. The outstanding difficulty of this problem makes it strongly advisable to employ a decomposition method, possibly coupled with a parallelisation of the solution process. However, existing decomposition methods are not well suited to solve the MCIS problem: some lead to a poor balance between subproblems, while others, like tree decomposition, are downright inapplicable. To enable the structural decomposition of such problems, Chmeiss et al. proposed an approach, TR-decomposition, acting at a low level: the microstructure of the problem. This approach had yet to be applied to the MCIS problem. We evaluate it in this context, aiming at reducing the size of the search space while also enabling parallelisation. The second problem that caught our interest is the sum colouring problem. It is an NP-hard variant of the widely known classical graph colouring problem. As in most colouring problems, it basically consists in assigning colours to the vertices of a given graph while making sure no neighbour vertices use the same colour. In the sum colouring problem, however, each colour is associated with a weight. The objective is to minimise the sum of the weights of the colours used by every vertex. This leads to generally harder instances than the classical colouring problem, which simply requires to use as few colours as possible. Only a few exact methods have been proposed for this problem. Among them stand notably a constraint programming (CP) model, a branch and bound approach, as well as an integer linear programming (ILP) model. We led an in-depth investigation of CP's capabilities to solve the sum colouring problem, while also looking into ways to make it more efficient. Additionally, we evaluated a combination of integer linear programming and constraint programming, with the intention of conciliating the strong points of these highly complementary approaches. We took inspiration from the classical backtracking bounded by tree decomposition (BTD) approach. We employ a tree decomposition with a strictly bounded height. We then derive profit from the complementarity of our approaches by developing a portfolio approach, able to choose one of the considered approaches automatically by relying on a number of features extracted from each instance.
Notre objectif est d’évaluer et de rendre opérationnelle la décomposition de problèmes d’optimisation sous contraintes. Nous nous sommes intéressés à deux problèmes en particulier : le problème de la recherche d’un plus grand sous-graphe commun (MCIS), et le problème de somme coloration minimale (MSCP). Il s’agit de problèmes NP-difficiles pour lesquels les approches de résolution complètes passent difficilement à l’échelle, et nous proposons de les améliorer à cet égard en décomposant ces problèmes en sous-problèmes indépendants. Les décompositions que nous proposons s’appuient sur la structure du problème initial pour créer des sous-problèmes de tailles équilibrées. Pour le MCIS, nous introduisons une décomposition basée sur la structure du graphe de compatibilité, et nous montrons que cette décomposition permet d’obtenir des sous-problèmes plus équilibrés que la méthode EPS classiquement utilisée pour paralléliser la résolution de problèmes en programmation par contraintes. Pour le MSCP, nous introduisons une nouvelle décomposition arborescente de hauteur bornée, et nous montrons comment tirer partie de la complémentarité de la programmation par contraintes et de la programmation linéaire en nombres entiers pour obtenir et résoudre les sous-problèmes indépendants qui en découlent. Nous proposons également une approche portfolio qui utilise des techniques d’apprentissage automatique pour choisir dynamiquement l’approche la plus performante en fonction du problème à résoudre.
Fichier principal
Vignette du fichier
thesis_minot_mael.pdf (987.25 Ko) Télécharger le fichier
Origin : Version validated by the jury (STAR)

Dates and versions

tel-01673531 , version 1 (30-12-2017)


  • HAL Id : tel-01673531 , version 1


Maël Minot. Investigating decomposition methods for the maximum common subgraph and sum colouring problems. Other [cs.OH]. Université de Lyon, 2017. English. ⟨NNT : 2017LYSEI120⟩. ⟨tel-01673531⟩
247 View
213 Download


Gmail Facebook Twitter LinkedIn More