Decomposition in large-scale mixed-integer optimization

Document Sample
Decomposition in large-scale mixed-integer optimization Powered By Docstoc
					Decomposition in large-scale mixed-integer optimization: applications in
                         production planning

                             Arie Quist (OM Partners)

Abstract:

One of the challenges in many industrial production planning environments is to
minimize the number of product-switches. Typically a product change incurs a cost or
time-loss due to cleaning, machine calibration or other human intervention. Mixed-
integer models of such campaign planning problems usually are far too large to be
solved to optimality, and the sub-optimal solutions can be improved by local search
techniques or sometimes even by visual inspection. By using a divide-and-conquer
approach, based on Dantzig-Wolfe decomposition, the MIP problems to be solved are
much smaller, leading to solutions that have much fewer product changes.

OM Partners is a software and consulting company delivering Supply Chain Planning
Solutions. OM Partners' offers a comprehensive software suite covering supply chain
processes related to network design, master planning, scheduling as well as execution.
The incorporation of solvers as well as integration modules on all planning levels
allows delivering highly interactive, intelligent and integrated systems.

With more than 150 clients and 300 implementations in Flow Shop (Corrugated and
Solid Board, Paper, Metals, Plastic, Textiles, Glass, ...) and Semi Process (Chemicals,
Pharmaceutics, Food & Beverages, Consumer Products, ...), OM Partners has
established solid partnerships with clients all over Europe and the USA for more than
15 years now. OM Partners has a long history in optimization and has been one of the
first companies offering a high-level modeling language for LP and MIP optimization
models.

In the first part of the presentation, we will present our company, with a focus on the
optimization techniques that are used in the various planning tools. In the second part,
we will describe the decomposition approach that is developed for campaign
planning. We show that application of the decomposition is very flexible using a
specialized link between our high-level modeling language and the MIP solver, such
that no specialized algorithmic knowledge is needed for the modeler. A demonstration
and some results show the viability of this approach.

The work on decomposition is a joint project with prof. Zeger Degraeve from London
Business School / Katholieke Universiteit Leuven.