SCIP-SDP is a plugin for SCIP to solve mixed integer semidefinite programs (MISDPs), i.e., semidefinite programs (SDPs) in which some variables are required to be integral.
It combines the branch-and-bound framework of SCIP with interior-point SDP-solvers to solve MISDPs using either a nonlinear branch-and-bound approach or an outer-approximation-based cutting-plane approach using linear programs (LPs). In addition to providing a constraint handler for SDP-constraints and a relaxator to solve continuous SDP-relaxations using interior-point solvers, SCIP-SDP adds several heuristics and propagators to SCIP.
The MISDPs can be read in using either the CBF-format or an extended
SDPA-format with support for integrality as well as rank-1
constraints. For a description of the extended SDPA-format see the
file sdpa_format.txt. The CBF-format is supported up to version 2,
see https://cblib.zib.de, and has also been extended to support rank-1
constraints.
To use the nonlinear branch-and-bound approach, one of the following SDP-solvers needs to be installed: DSDP, SDPA, or MOSEK. For more information about the installation of SCIP-SDP see the INSTALL file.
SCIP-SDP can read MISDPs in CBF or extended SDPA format.
SCIP-SDP contains presolving and propagation methods as well as primal heuristics.
SCIP-SDP supports rank-1 constraints by adding quadratic constraints for each 2 by 2 minor. Such problems are usually (very) hard to solve.
One can switch between solving SDPs or using an outer approximation via LP-relaxations.
SCIP-SDP extends SCIP and can be incorporated into other codes similar to how this can be done for SCIP.
SCIP-SDP features many parameters to determine its behavior and it inherits all SCIP parameters. Here, we highlight some:
misc/solvesdps: Determines whether SDPs (1) or LPs (0) are
solved. It depends on the instance which approach is faster.
relaxing/SDP/sdpsolverthreads sets the number of threads used for
solving SDPs. By default it is set to 1, since this seems to be the
fastest for most instances. For larger SDPs it might help to
increase this number, where a value of -1 corresponds to an
automatic choice, if this is supported by the SDP solver.
Tolerances: numerics/feastol and numerics/dualfeastol (default
1e-5) are SCIP parameters that determine the feasibility
tolerances. Note that the default is a bit weaker than the default
SCIP parameter values (1e-6 and 1e-7, respectively).
relaxing/SDP/sdpsolverfeastol and relaxing/SDP/sdpsolvergaptol
(default 1e-5) determine the tolerances used for the SDP-solvers.
For Mosek the feasibility tolerance is always tightened by 0.1,
because this generated more reliable results. Depending on the
instance, changing these parameters might have a dramatic effect on
performance and correctness of the results (due to numerical
issues).
Slater condition: The solution process of interior-point methods for
SDPs depends on the Slater condition. One can determine the Slater
condition for the primal and dual problem by changing
relaxing/SDP/slatercheck (0: no, 1: yes but only for statistics, 2:
yes and print warning for every problem not satisfying primal and
dual Slater condition).
A doxygen documentation is given at https://wwwopt.mathematik.tu-darmstadt.de/scipsdp/ .
Reports for the SCIP Optimization Suites listed at https://www.scipopt.org/index.php#cite each contain a part on the developments in SCIP-SDP.
Many methods used in SCIP-SDP are described in the following dissertations:
Sonja Mars (2013) "Mixed-Integer Semidefinite Programming with an Application to Truss Topology Design"
Tristan Gally (2019) "Computational Mixed-Integer Semidefinite Programming"
Frederic Matter (2022) "Sparse Recovery Under Side Constraints Using Null Space Properties"
Copyright:
Licensed under the Apache License, Version 2.0 (the "License"), see the LICENSE file.
You may not use this file except in compliance with the License.
You may obtain a copy of the License at https://www.apache.org/licenses/LICENSE-2.0 .
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.