INFORMS 2021 Program Book

INFORMS Anaheim 2021

ME30

3 - An Overview of Learning Enabled Stochastic Optimization Suvrajeet Sen, Univ. of Southern California, Santa Monica, CA, 90403, United States The combination of learning and optimization provides the scientific basis for analytics, allowing rapid and defensible data driven decisions. We outline a broad agenda which allows new modeling and algorithmic tools to be integrated in one framework. ME29 CC Room 207C In Person: OR/ML Practice at Amazon General Session Chair: Kaiyue (Kay) Zheng, Amazon, Bellevue, WA, 98004, United States 1 - Optimal Trailer Management Using a Multi-Period Multi-Location Variant on the Newsvendor Model Dmitriy Belyi, Amazon.com, Austin, TX, 78750-4024, United States We consider the trailer management problem in Amazon’s transportation network, where we must allocate available trailers to accommodate all transportation needs at the lowest cost. This problem is fraught with uncertainty and high dimensionality, and is difficult to model and solve using standard approaches. In this talk, we consider a novel modeling approach for this problem, reminiscent of the classic newsvendor. We also present an efficient parallelized solution methodology based on marginal costs. 2 - Sortation Optimization and Allocation Planning in Amazon Middle Mile Network Kay Zheng, Research Scientist, Amazon, Bellevue, WA, 98004, United States Middle mile network involves the transportation and sortation of goods from warehouses to the last mile carriers that provides final-mile delivery services. Amazon middle mile transportation paths can include one or more sortation processes along the way and span hundreds or even thousands of miles. Sortation resource optimization and allocation at the middle mile network plans for where, when and how to sort packages along their shipment paths so as to achieve cost, speed or other goals given limited resources at warehouses and on the road. In this work, we present a mixed-integer programming based planning approach to model the trade-offs in the network and discuss the complexities due to the interconnectedness of network structure. In Person: Applying for SBIR/STTR Grants/Data Analytics for Structured Data from Advanced Sensing Systems General Session Chair: Arman Sabbaghi, Purdue University, West Lafayette, IN, 47907- 2067, United States Co-Chair: Andi Wang, Georgia Institute of Technology, Atlanta, GA, 30318-5546, United States 1 - Applying for SBIR/STTR Grants Arman Sabbaghi, Purdue University, West Lafayette, IN, 47907- 2067, United States The Small Business Innovation Research (SBIR) and Small Business Technology Transfer (STTR) programs provide exciting opportunities for researchers who have created domestic small businesses to engage in federal research/research and development, with the potential for commercialization. The panelists in this session will discuss SBIR/STTR programs and share information about applying to these programs. 2 - Wavelet Basis Function for Meltpool Monitoring in AM Siqi Zhang, The Pennsylvania State University, State College, PA, United States The characteristics of melt pools are critical for process monitoring and control in AM process. However, there are practical issues pertinent to in-situ monitoring of melt-pool characteristics (e.g., a large volume of time-varying melt-pool imaging data, sensitivity to many process parameters (i.e., laser power)). Hence, this paper presents a parametric approach of wavelet basis function to model and monitor melt-pool variations in AM. Specifically, we designed a sparse kernel-weighted regression model to represent the high-dimension imaging data. Experimental results on real-world data demonstrated the effectiveness of wavelet basis functions to represent and monitor the melt-pool imaging data in AM. ME30 CC Room 207D

ME26 CC Room 206A In Person: Packing and Scheduling General Session Chair: Christopher Muir, Georgia Institute of Technology, Atlanta, GA, United States 1 - Incremental Packing Problems Lingyi Zhang, Columbia University, New York, NY, 10027-6702, United States, Yuri Faenza, Danny Segev Incremental packing problems extend classical packing problems (like knapsack, stable set, generalized assignment, etc.) to a multi-period setting. In each period, we are allowed to pack a previously unpacked item (if there is enough capacity available), but are not allowed to unpack a previously packed item. In this talk, we present formulations and approximation algorithms for incremental packing problems via a reduction to a multi-machine scheduling problem, and we moreover investigate limits of using single-period techniques in the multi-period setting. 2 - Capacity Expansion in the College Admission Problem Federico Bobbio, PhD Candidate, University of Montreal, Montreal, QC, Canada, Margarida Carvalho, Alfredo Torrico, Andrea Lodi The college admission problem plays an important role in several real-world allocation mechanisms. In particular, the student-oriented deferred acceptance algorithm is known to produce a stable matching that is weakly preferred by every student. However, if an extra position is available in a single university, then a subset of the students would improve their allocations. This raises a natural question: To which universities should we allocate B extra positions in order to produce the best possible allocation for all the students? In this work, we study the expansion of the capacities in the college admission problem with strict and complete preferences. Our main contribution is twofold: First, we provide a theoretical understanding of the underlying mathematical structure of the problem; second, we propose an algorithmic approach to solve the problem. 3 - Submodular Interval Scheduling Christopher Muir, Georgia Institute of Technology, Atlanta, GA, 30318, United States, Alejandro Toriello Given a set of jobs with fixed start/end times, fixed interval scheduling asks the decision maker to group the jobs into schedules such that jobs within a schedule do not overlap in time. The decision maker wants to minimize the sum of schedule costs. Submodular cost functions are commonly used as they model economies of scale; however, interval scheduling with submodular costs is NP- Hard. We propose a branch-and-price approach for submodular fixed interval scheduling. For some specific cost functions, we present efficient algorithms for solving the linear relaxation of the underlying column-generation model. We evaluate our approach using random instances and instances derived from cloud services applications. We show that branch-and-price is able to efficiently solve large instances, outperforming standard integer programming models when available. ME28 CC Room 207B In Person: Data Science and Stochastic Optimization General Session Chair: Suvrajeet Sen, Univ. of Southern California, Santa Monica, CA, 90403, United States 1 - Nonparametric Stochastic Decomposition Shuotao Diao, University of Southern California, Los Angeles, CA, 90007-2490, United States, Suvrajeet Sen We study the mathematical fusion of non-parametric estimation and stochastic decomposition (SD) algorithms which we refer to as Non-parametric SD. This permits simultaneous updates of the expected value objective, as well as first- stage decisions using k nearest neighbor (kNN) estimation to calculate a new minorant of the current kNN benchmark function. Both convergence and computational results will be presented. 2 - Primal-dual Incremental Gradient Method Fornonsmooth and Convex Optimization Problems Afrooz Jalilzadeh, The University of Arizona, State College, PA, 16801-4415, United States n this talk, we consider a nonsmooth convex finite-sum problemwith a conic constraint. To overcome the challenge of projecting onto the con-straint set and computing the full (sub)gradient, we introduce a primal-dualincremental gradient scheme where only a component function and two con-straints are used to update each primal-dual sub-iteration in a cyclic order.We demonstrate an asymptotic sublinear rate of convergence in terms of sub-optimality and infeasibility which is an improvement over the state-of-the-artincremental gradient schemes in this setting.

95