MIP Formulations for the Steiner Forest Problem

TL;DR

This paper compares existing integer programming formulations for the Steiner Forest problem, introduces three new formulations, and demonstrates that these new models provide tighter bounds and improved computational performance.

Contribution

The paper proposes three new integer programming formulations for the Steiner Forest problem that yield better bounds and computational efficiency than existing models.

Findings

New formulations achieve bounds within a factor of 2 of the integer optimum.

Experimental results show the new models outperform existing formulations.

Bounds are significantly tighter than previous tractable models.

Abstract

The Steiner Forest problem is among the fundamental network design problems. Finding tight linear programming bounds for the problem is the key for both fast Branch-and-Bound algorithms and good primal-dual approximations. On the theoretical side, the best known bound can be obtained from an integer program [KLSv08]. It guarantees a value that is a (2-eps)-approximation of the integer optimum. On the practical side, bounds from a mixed integer program by Magnanti and Raghavan [MR05] are very close to the integer optimum in computational experiments, but the size of the model limits its practical usefulness. We compare a number of known integer programming formulations for the problem and propose three new formulations. We can show that the bounds from our two new cut-based formulations for the problem are within a factor of 2 of the integer optimum. In our experiments, the formulations prove to be both tractable and provide better bounds than all other tractable formulations. In particular, the factor to the integer optimum is much better than 2 in the experiments.