2016 INFORMS Annual Meeting Program

SA15

INFORMS Nashville – 2016

SA16 105A-MCC Inverse Optimization: Theory Sponsored: Optimization, Optimization Under Uncertainty Sponsored Session Chair: Taewoo Lee, Rice University, #217, 7010 Staffordshire Street, Houston, TX, 77030, United States, taewoo.lee@utoronto.ca Co-Chair: Timothy C.Y. Chan, University of Toronto, Toronto, ON, Canada, tcychan@mie.utoronto.ca 1 - Goodness-of-fit In Multi-point Inverse Optimization Optimization Rafid Mahmood, University of Toronto, Toronto, ON, Canada, rafid.mahmood@mail.utoronto.ca, Timothy Chan, Taewoo Lee, Daria Terekhov Inverse optimization is a model fitting technique that uses observed points to impute the cost function of an unknown optimization problem. Applications of inverse optimization often rely on ad-hoc or informal methods to evaluate the fit quality of the inverse solution to the data. A previous work introduced a general formulation for inverse optimization with a single observation and a measure for the goodness-of-fit. We extend both of these results to the case of multiple observed points. Our techniques are capable of comparing different models and identifying outliers that do not fit well with the remaining points. 2 - Inverse Optimization For Determining Constraint Parameters Neal Kaw, University of Toronto, Toronto, ON, Canada, neal.kaw@mail.utoronto.ca, Timothy Chan Most inverse optimization literature has focused on determining the objective function of an optimization problem, given an observed solution. In this work, we develop inverse optimization models that additionally determine unspecified parameters of the feasible set. First, we propose an inverse linear programming model to determine all problem data. Second, we propose inverse robust linear programming models to determine a cost vector and unspecified parameters of the uncertainty set, for two types of uncertainty: interval uncertainty and cardinality constrained uncertainty. 3 - Robust Inverse Optimization Kimia Ghobadi, MIT, Cambridge, MA, United States, kimiag@mit.edu, Daria Terekhov, Houra Mahmoudzadeh, Taewoo Lee In this talk, we explore the robustification of inverse optimization. Our work is motivated by problems in which the observation of the solution is partial, noisy, or uncertain. We build an uncertainty set around the observation and derive an inverse model that finds a cost vector that protects against the worst case scenario in the given uncertainty set. Our model generalizes previous work on single- observation inverse models. It can also be seen as more general than inverse optimization with multiple points, since the points can be thought of as a sample from the uncertainty set. SA17 105B-MCC Optimal Statistical Learning Sponsored: Optimization, Optimization Under Uncertainty Sponsored Session Chair: Nana Kwabena Aboagye, Princeton University (ORFE), 1 Nassau Hall, Princeton, NJ, 08544, United States, aboagye@princeton.edu 1 - Uncertain Date Envelopement Analysis Allen Holder, Rose-Hulman Institute of Technology, holder@rose-hulman.edu We motivate an inverse optimization problem that calculates a decision making unit’s maximum efficiency within the context of uncertain data envelope analysis. One of the sub-problems is a robust linear program, but unlike a traditional robust model that sacrifices the objective to hedge against uncertainty, the data envelope model leverages uncertainty to promote efficiency. We apply the method to a set of prostate radiotherapy treatments to help discern appropriate treatments. 2 - Optimal Learning Of Expensive Quadratic Functions Nana Kwabena Aboagye, Princeton University, aboagye@princeton.edu We study the problem of learning the unknown parameters of an expensive function where the true underlying surface can be described by a quadratic polynomial. We present a previously studied Bayesian optimization algorithm known as the knowledge gradient for the parametric belief model. Originally established in the limited context of drug discovery (see Negoescu et al. (2011), the knowledge gradient for the parametric belief model remains under-studied with regards to its behavior. We seek to understand the behavior of this algorithm and exploit this understanding to derive a simple heuristic that performs just as well as the knowledge gradient for the parametric belief model.

3 - Response Surface Methodology In Plant Breeding Reka Howard, University of Nebraska – Lincoln, NE, rekahoward@gmail.com, William Beavis, Alicia Carriquiry We introduce Response Surface Methodology (RSM) as a strategy to find the combination of attribute levels that results in accurate predictions for a given genomic prediction (GP) method, and compare GP methods. We illustrate RSM with a simulated example where the response we optimize is the difference between prediction accuracy using the parametric best linear unbiased prediction (BLUP) and the nonparametric support vector machine (SVM). The greatest impact on the response is due to the genetic architecture of the population and the heritability. When epistasis and heritability are highest, the advantage of using the SVM versus the BLUP is greatest. 4 - A New Genomic Selection Approach Lizhi Wang, Iowa State University, lzwang@iastate.edu, Matthew Goiffon, Guiping Hu, Aaron Kusmec, Patrick Schnable Conventionally, plant breeders make selection decisions based on phenotype observations and intuitive judgement. The advent of genotyping techniques provides breeders with much more informative genomic data. However, the enormous volume and complexity of the genomic data also present great challenges in extracting the useful information deeply buried in the mountains of data. We present a new approach for genomic selection and demonstrate its improvement over previous methods using computer simulation with realistic genomic data. Big Data in the E-Commerce Deliveries Invited: Modeling and Methodologies in Big Data Invited Session Chair: Chung-Yee Lee, HKUST, IELM Dept. HKUST, Clear Water Bay, Hong Kong, 0000, Hong Kong, cylee@ust.hk 1 - The Benefits Of Randomization In Warehousing And Logistics John Carlsson, University of Southern California, jcarlsso@usc.edu A recent innovation in warehousing and logistics has been the use of randomization, such as Amazon’s random stow, in which warehouse items are scattered throughout the floor map as opposed to being concentrated in one area. We use a continuous approximation model to describe how such a policy is beneficial in the long run. 2 - Resource Allocation With Unmanned Aerial Vehicle Siyuan Song, University of Southern California, siyuanso@usc.edu, John Gunnar Carlsson Unmanned aerial vehicles, commonly known as drones, have become more widely utilized in delivery nowadays. We study the efficiency of a so-called ‘horsefly’ delivery system, in which drones are used in conjunction with truck. We propose a mathematical formulation of a ‘horsefly’ problem followed by some general properties of optimal solutions. Then some approximation results, including an approximation algorithm, are given to illustrate the benefit of horsefly system on a large scale. Lastly, we compare some practical heuristic algorithms in different scenarios for best choice in each case. 3 - The Last Mile Rush Song Zheng, Cainiao Network, Hang Zhou, China, zhengsong.zs@alibaba-inc.com, Lijun Zhu As e-commerce keeps its impressive growth, a large percentage of express orders are generated by e-commerce. In China, for example, it is over 60 percent. Increasing investments rush into China’s express delivery industry, which now has thousands of delivery companies and millions of delivery workers. Alibaba group and Cainiao Network are building China Smart Logistics Network and developing a huge ecosystem with all major logistics companies in China. We will present an optimal solution to the last mile delivery, more specifically, arranging thousands of courier to delivery all kinds of packages in cities including online e- commerce packages and offline O2O packages. SA15 104E-MCC

22