2015 Informs Annual Meeting

WD14

INFORMS Philadelphia – 2015

WD15 15-Franklin 5, Marriott Optimization Methodology I Contributed Session Chair: David Valembois, Head Of Business Development, ENGIE, Avenue Enstein, 2a, Louvain-la-Neuve, 1300, Belgium, david.valembois@external.engie.com 1 - Globally Solving Non-Convex Quadratic Programming Problems via Linear Integer Programming Techniques Wei Xia, Lehigh University, 312 Brodhead Ave., Bethlehem, PA, 18015, United States of America, wex213@lehigh.edu, Luis Zuluaga, Juan Vera Quadratic programming (QP) problems are NP-hard optimization problems. We propose an alternative way of globally solving nonconvex QPs by exploiting of the Karush-Kuhn-Tucker conditions to linearize its objective. Then we reformulate the problem as a mixed integer linear problem with binary variables to model the KKT complementary constraints. We compare the performance of this solution approach with the benchmark global QP solver QuadprogBB on a variety of test QP instances. 2 - An Engie Novel Approach to the Solution of Thermal Unit Commitment Problem with Coupling Constraints Melodie Mouffe, ENGIE, Avenue Einstein, 2A, NewTech Center, Louvain-la-Neuve, 1348, Belgium, melodie.mouffe@engie.com, Dimitri Tomanos Engie developed PoweredPegase to optimize the commitment of its large number of plants on a medium term horizon, while taking into account supply delivery, storage and ancillary services. Getting to find a solution using a fine time granularity is difficult to achieve, specifically when dealing with global constraints. We propose a novel approach to find an accurate solution in a reasonable amount of time, without performing any model approximation, implementing a “smart” time decomposition. 3 - Poweredpegase : Engie’s Next-gen Software for Joint Gas and Power Portfolio MT Optimization

3 - Dynamic Allocations for Cooperative Games under Uncertainty with Risk-averse Players Nelson Uhan, Assistant Professor, United States Naval Academy, Mathematics Department, Chauvenet Hall, Annapolis, MD, 21402, United States of America, uhan@usna.edu, Alejandro Toriello We consider a class of cooperative games in which the costs of cooperation are uncertain and evolve over time, and the players are risk averse. These games generalize the classic linear production game, and as a result, model a variety of cooperative settings. We give sufficient conditions for the existence of an allocation in the strong sequential core - the set of allocations that distribute costs as they are incurred and are stable against coalitional defections at any point in time. 4 - Ambiguous Stochastic Programs with Variation Distance Hamed Rahimian, PhD Student, The Ohio State University, Integrated Systems Engineering, Columbus, OH, 43210, United States of America, rahimian.1@osu.edu, Guzin Bayraksan, Tito Homem-de-mello Ambiguous stochastic programs relax the assumption of known distributions in stochastic programming and instead use an ambiguity set of distributions. We focus on the variation distance to form the ambiguity set, examine the resulting model properties, and propose a decomposition-based algorithm to solve it. We characterize a minimal scenario tree, where the presence of every scenario is critical in determining the optimal objective function. Data Driven Optimization and Applications I Sponsor: Optimization/Optimization Under Uncertainty Sponsored Session Chair: Michael Kim, University of Toronto, 5 King’s College Road, Toronto, Canada, mikekim@mie.utoronto.ca 1 - Robust Optimization with Learning Andrew Lim, National University of Singapore/Department of Decision Sciences, Mochtar Riady Building, BIZ1 08-69, 15 Kent Ridge Drive, Singapore, Singapore, andrewlim@nus.edu.sg, Michael Kim We consider a robust optimization problem with learning in the setting of a Bayesian mixture model and show how it is a natural framework for modeling customer heterogeneity in business analytics applications. Asymptotic equivalence to a mean-variance problem is established. More generally, we illustrate how robust Bayesian models are a natural framework for combining concerns for robustness and learning about uncertainty sets. 2 - Coordinating Pricing and Inventory Replenishment with Nonparametric Demand Learning Boxiao (Beryl) Chen, University of Michigan-Ann Arbor, 1205 Beal Avenue, Ann Arbor, MI, 48109, United States of America, boxchen@umich.edu, Xiuli Chao, Hyun-Soo Ahn We consider a finite horizon problem of dynamic pricing and inventory control for a nonperishable product. We develop a data-driven policy that does not require explicit information about the demand distribution and show that it is asymptotically optimal and converges at the fastest possible speed. 3 - Data-driven Assortment Optimization Velibor Misic, Massachusetts Institute of Technology, 77 Massachusetts Avenue, E40-147, Cambridge, MA, 02139, United States of America, vvmisic@mit.edu, Dimitris Bertsimas We present a practical method for transforming limited historical transaction data into effective assortment decisions. Our method consists of estimating a choice model from the data by efficiently solving a large-scale linear program, and finding the optimal assortment by solving a tractable mixed-integer program. Our method leads to assortments that achieve near-optimal revenues and significantly outperform assortments derived from other parametric and nonparametric approaches. 4 - Geo-demand Estimation and Inventory Allocation in Online Retailing Long He, University of California, Berkeley, 1117 Etcheverry Hall, Berkeley, CA, 94720, United States of America, longhe@berkeley.edu, Zhiwei (Tony) Qin For existing items, we first formulate the multi-dimensional geo-demand distribution estimation problem for existing items as a robust low-rank tensor recovery problem in a convex optimization framework. We also propose a tailored algorithm based on the alternating direction augmented Lagrangian method. For new items, we develop the estimation based on nested logit model and discuss the inventory allocation policy in practice. WD14 14-Franklin 4, Marriott

David Valembois, Head Of Business Development, ENGIE, Avenue Enstein, 2a, Louvain-la-Neuve, 1300, Belgium, david.valembois@external.engie.com

PoweredPegase is ENGIE’s next-gen software for joint gas & power portfolio MT optimization. PP delivers reliable and accurate solutions in remarkable short time thanks to disruptive “smart time” approach. PP open “working modules” allow modifying models by integrating your specific context. If flexibility and accuracy really matter, PoweredPegase is the answer. ENGIE (GDF-Suez) is the No. 1 Independent Energy Producer in the World. Winter is Coming–Optimize Your Energy with PoweredPegase.

WD16 16-Franklin 6, Marriott

Game Theory V Contributed Session Shuo Zeng,University of Arizona, McClelland Hall 430, 1130 E. Helen Street, Tucson AZ 85721, United States of America, shuozeng@email.arizona.edu 1 - Strategic Delay in Networked Bargaining Thanh Nguyen, Krannert School of Management, Purdue University, West Lafayette, IN, United States of America, nguye161@purdue.edu, Vijay Subramanian, Randall Berry We study decentralized markets involving producers and consumers that are facilitated by middlemen. We do this by analyzing a non-cooperative networked bargaining game. We show that sunk cost problems and a heterogeneous network can give rise to delay or failure in negotiation, and therefore, reduce the total trade capacity of the network. In the limiting regime of extremely patient agents, we provide a sharp characterization of the trade pattern and the segmentation of these markets. 2 - A Joint Work on Knowledge Management and Operations

Management: How to Keep Firm-Specific Knowledge Kai Luo, Assistant Professor, KEDGE Business School, 450 Avenue François Arago CS 90262, La Garde, France, luokailk@sohu.com, Salomee Ruel, Sajjad Jasimuddin

Firm-specific knowledge (FSK) acquisition and maintenance is essential for a firm to fill knowledge gaps and enhance its competitive advantage. We start with qualitative research that identifies the existence and important features of the FSK problem, followed by quantitative research formulated as a principal-agent problem, providing closed-form solutions. We used our model to compare French and American firm styles via 40 numerical examples and develop managerial insights.

458