A heuristic use of dynamic programming to upperbound treewidth

TL;DR

This paper introduces a heuristic algorithm that uses dynamic programming on potential maximal cliques to efficiently compute upper bounds for treewidth, significantly outperforming previous heuristics.

Contribution

It presents a novel iterative improvement method for treewidth upper bounds based on potential maximal cliques, leveraging a dynamic programming approach.

Findings

Algorithm vastly outperforms previous heuristics

Efficient computation of treewidth upper bounds

Effective use of potential maximal cliques in heuristics

Abstract

For a graph $G$, let $\Pi(G)$ denote the set of all potential maximal cliques of $G$. For each subset $\Pi$ of $\Pi(G)$, let $\tw(G, \Pi)$ denote the smallest $k$ such that there is a tree-decomposition of $G$ of width $k$ whose bags all belong to $\Pi$. Bouchitt\'{e} and Todinca observed in 2001 that $\tw(G, \Pi(G))$ is exactly the treewidth of $G$ and developed a dynamic programming algorithm to compute it. Indeed, their algorithm can readily be applied to an arbitrary non-empty subset $\Pi$ of $\Pi(G)$ and computes $\tw(G, \Pi)$, or reports that it is undefined, in time $|\Pi||V(G)|^{O(1)}$. This efficient tool for computing $\tw(G, \Pi)$ allows us to conceive of an iterative improvement procedure for treewidth upper bounds which maintains, as the current solution, a set of potential maximal cliques rather than a tree-decomposition. We design and implement an algorithm along this approach. Experiments show that our algorithm vastly outperforms previously implemented heuristic algorithms for treewidth.