Sparsity of Lift-and-Project Cutting Planes
It is well-known that sparsity (i.e. having only a few nonzero coefficients) is a desirable property for cutting planes in mixed-integer programming. We show that on the MIPLIB 2003 problem instance set, using only 10 very dense cutting planes (compared to thousands of constraints in a model), leads to a run time increase of 25 % on average for the LP-solver. We introduce the concept of dual sparsity (a property of the row-multipliers of the cut) and show a strong correlation between dual and primal (the usual) sparsity. Lift-and-project cuts crucially depend on the choice of a so-called normalization, of which we compared several known ones with respect to their actual and possible sparsity. Then a new normalization is tested that improves the dual (and hence the primal) sparsity of the generated cuts.
I owe many thanks to my internship advisor Laci Ladanyi in the CPLEX group at IBM as well as my university advisor Volker Kaibel. Additionally, I want to thank Andrea Lodi, Tobias Achterberg and Roland Wunderling for several stimulating discussions.
- 3.Balas, E., Ceria, S., Cornuéjols, G.: A lift-and-project cutting plane algorithm for mixed 0–1 programs. Math. Program. 58, 295324 (1993)Google Scholar
- 5.Fischetti, M., Lodi, A., Tramontani, A.: On the separation of disjunctive cuts. Math. Program. 128, 205–230 (June 2011)Google Scholar
- 6.Wolter, K.: Implementation of cutting plane separators for mixed integer programs. Masters thesis (2006)Google Scholar