Section: Partnerships and Cooperations

International Initiatives

SAMBA

• Title: Combinatorial optimization problems

• International Partner (Institution - Laboratory - Researcher):

  • Pontifícia Universidade Católica do Rio de Janeiro, Brazil

  • Universidade Federal Fluminense (UFF), Brazil

  • Universidad Adolfo Ibañez, Chile

• Duration: 2014 - 2017

• See also: https://wiki.bordeaux.inria.fr/realopt/pmwiki.php/Project/Samba

The renewed project builds on our previous SAMBA output with new emphasis on 4 axis:

1. Algorithmic Performance Enhancements: In the line of the considerable algorithmic speed-up that we obtained recently in SAMBA by developping stabilization techniques, warm-starting techniques (with memorized basis to initialize the node of the enumeration tree), and strong branching techniques (that limit the size of the enumeration tree), we aim to develop intensive preprocessing techniques building on contraint propagation. Further contibutions shall consist in integrating dynamic aggregation-disagregation techniques.

2. Extending the Dantzig-Wolfe reformulation paradigm. The current SAMBA project has lead to finalizing a technique called “column generation for extended formulations” which can be understood as a generalization of Dantzig-Wolfe reformulation: To favour early convergence, the Dantzig-Wolfe reformulation is lifted into an extended variable space where the recombination of solutions arises. Further extension is built in the proposal of Goycoolea et al.

3. Combining Dantzig-Wolfe decomposition with Benders': In a stochastic environement, a numerically realistic approach in to build solutions that resists to worst case perturbations drawn within a contrained uncertainty set. In such context, bilevel optimization naturally arises: the second level models the worst case reaction of the system, along with our recourse, considering as fixed, the decisions of the first level of optimization. The model constraints are therefore decomposed into first level and second level, suited for Benders approach. When the first stage is a multiple resource planning applications, a strong model leading to good continuous approximation can be obtained by reformulating the problem in terms of variables that encode a work allocation for an individual resource (this is known as the Dantzig-Wolfe decomposition approach).

4. Build-up our BAPCOD software platform for new benchmarks and industrial transfer: the aim is to translate our research output into efficient code, to develop high level interface that free the end users from the expert knowledge normally required for complex decomposition based solution.