Combinatorial optimization under ellipsoidal uncertainty
| dc.contributor.advisor | Buchheim, Christoph | |
| dc.contributor.author | Ilyina, Anna | |
| dc.contributor.referee | Mutzel, Petra | |
| dc.date.accepted | 2017-06-28 | |
| dc.date.accessioned | 2017-09-11T06:24:21Z | |
| dc.date.available | 2017-09-11T06:24:21Z | |
| dc.date.issued | 2017 | |
| dc.description.abstract | We study combinatorial problems with ellipsoidal uncertainty in the objective function concerning their theoretical and practical solvability. Ellipsoidal uncertainty is a natural model when the coefficients are normally distributed random variables. Robust versions of typical combinatorial problems can be very hard to solve compared to their linear versions. Complexity and approaches differ fundamentally depending on whether uncorrelated or correlated uncertainty occurs. We distinguish between these two cases and consider first the unconstrained binary optimization under uncorrelated ellipsoidal uncertainty. For this we develop an algorithm which computes an optimal solution by merely sorting the variables and, correspondingly, has a running time of O(n log n). The algorithm is based on the diminishing returns-property, which is characteristic for submodular functions. We introduce a new and a more general p-norm-uncertainty and show that with only slight modifications the sorting algorithm can be easily applied. We also extend the algorithm to general integer variables, which in this case only leads to a pseudo-polynomial time. The next step to the general case is investigation of problems with arbitrary combinatorial sets X ⊆ {0, 1}n under uncorrelated ellipsoidal uncertainty. For this case we embed the O(n log n)-algorithm for the unconstrained binary problems into a Lagrangean decomposition approach. The approach separates the objective function from the combinatorial structure applying Lagrangean relaxation to some artificial connecting constraints. This creates two subproblems, one of which is the linear version of the combinatorial problem and the other one is just the unconstrained binary uncorrelated problem, which can be solved using the O(n log n)-algorithm. The solutions of the subproblems are used to obtain primal and dual bounds which are used in a branch and bound-approach. The approach shows an excellent performance in practice. In the correlated case already the unconstrained binary problem turns out to be strongly NP-hard. Here we also define a branch and bound-approach, now with lower bounds determined by underestimation of the given ellipsoid with certainly defined axis-parallel ellipsoids. We use this idea to extend the decomposition approach to general combinatorial problems under correlated uncertainty. In contrast to the uncorrelated case the uncertain subproblem of the decomposition is here strongly NP-hard in itself. We solve it approximately using the developed underestimators which are determined in a preprocessing step. The approach offers room for improvement concerning in the primal extent a faster computation of the underestimators, which is done by solving semidefinite programs. | en |
| dc.identifier.uri | http://hdl.handle.net/2003/36086 | |
| dc.identifier.uri | http://dx.doi.org/10.17877/DE290R-18102 | |
| dc.language.iso | en | de |
| dc.subject | Robuste Optimierung | de |
| dc.subject | Ellipsoidale Unsicherheit | de |
| dc.subject | Kombinatorische Optimierung | de |
| dc.subject.ddc | 510 | |
| dc.subject.rswk | Robuste Optimierung | de |
| dc.subject.rswk | Unsicherheit | de |
| dc.subject.rswk | Kombinatorische Optimierung | de |
| dc.title | Combinatorial optimization under ellipsoidal uncertainty | en |
| dc.type | Text | de |
| dc.type.publicationtype | doctoralThesis | de |
| dcterms.accessRights | open access |