Informs Annual Meeting 2017

TE64

INFORMS Houston – 2017

3 - Adaptive Secant Tangents Averaged Stochastic Approximation Marie Chau, Virginia Commonwealth University, 1015 Floyd Avenue, Richmond, VA, 23220, United States, mchau@vcu.edu, Jong J. Lee, Michael C. Fu We introduce Adaptive Secant Tangents AveRaged SPSA (aSTAR-SPSA) for multivariate stochastic optimization problems. Under the setting, where direct and indirect gradients are available, aSTAR-SPSA employs a hybrid Hessian, which is a weighted combination of gradient-free (GF) and gradient-based (GB) Hessians. GF consists of symmetric finite differences (SFD) of SPSA gradients, and the SFD of the associated direct gradients generate GB. We propose weights to minimize two criteria under the Frobenius norm. We prove convergence, establish asymptotic normality, and investigate the empirical performance of aSTAR-SPSA against second-order SA methods through numerical experiments. 4 - Integrating Lyapunov Optimization with Monte Carlo Simulation for Online Decisions Ashutosh Nayak, PhD Candidate, Purdue University, 45 N Salisbury Street, Apt 9, West Lafayette, IN, 47906, United States, nayak2@purdue.edu, Seokcheon Lee With the explosion of real-time information, online decision making is increasingly gaining popularity in different applications. Smart Microgrids receive dynamic load requests from consumers and time-varying electricity from renewables. The stochastic nature of demand and supply makes online load scheduling decisions in Microgrid imperative. Lyapunov optimization is an elegant and simple stochastic dual-gradient method that does not consider estimation of future information. In this work, we integrate Lyapunov optimization with Monte Carlo simulation to construct an online algorithm for dynamic load scheduling decisions based on historical decisions and future estimation. Chair: Changyue Song, PhD, University of Wisconsin-Madison, 1513 University Ave, Madison, WI, 53715, United States, csong39@wisc.edu Co-Chair: Kaibo Liu, UW-Madison, Madison, WI, 53706, United States, jacoblkb@gmail.com 1 - Manufacturing Functional Data Analysis Ran Jin, Virginia Tech, 1145 Perry Street, Durham Hall, Room 250 (0118), Blacksburg, VA, 24061, United States, jran5@vt.edu, Hongyue Sun Modern manufacturing system generates spatially and temporally dense data sets. This talk focuses on manufacturing modeling problems with functional data, where the models will be used in data-driven decision making in smart manufacturing. Examples in functional variable selections, in situ process modeling, and data interpretation from natural language processing perspective will be discussed. 2 - Monitoring for Changes in the Nature of Stochastic Textured Surfaces Anh Tuan Bui, Northwestern University, Technological Institute This work aims to monitor for global changes in the nature of stochastic textured surfaces, which are fundamentally different than the profiles and images that are the focus of most prior profile monitoring works. We represent normal in-control behavior by using supervised learning models to implicitly characterize the joint distribution of the surface image pixels. Based on this characterization, we develop a monitoring statistic using likelihood-ratio principles to detect changes in the surface nature, relative to the in-control nature. Our approach can detect general surface changes with no prior knowledge of the changes. We demonstrate the approach with a textile example. 3 - Early Detection Vessel Delays using Combined Historical and Real-time Information Sungil Kim, Assistant Professor, UNIST, Ulsan, Korea, Republic of, sungil.kim@unist.ac.kr Detecting vessel delays in advance or in real time is important in order to fulfill the expectations of customers and to help customers reduce delay costs. However, the early detection of vessel delays faces the challenges of numerous uncertainties, including weather conditions, port congestion, booking issues, and route selection. This paper proposes a data-driven method for the early detection of vessel delays: in our new framework of refined case-based reasoning, real-time S-AIS vessel tracking data are utilized in combination with historical shipping data. Real data examples from a logistics company demonstrate the effectiveness of the proposed method. 2145 Sheridan Road, Room C210, Evanston, IL, 60208, United States, atbui@u.northwestern.edu, Daniel Apley TE64 370E Data Analytics for System Improvement II Sponsored: Data Mining Sponsored Session

4 - A Generic Multisensor-based Degradation Model Changyue Song, University of Wisconsin-Madison, Rm 3221, Mechanical Engineering Building, 1513 University Ave, Madison, WI, 53715, United States, csong39@wisc.edu, Kaibo Liu Nowadays multiple sensors have been widely used to monitor the degradation status of a unit simultaneously, which enables an unprecedented opportunity for condition-based monitoring and maintenance. However, most of the existing studies focus on the modeling and prognosis based on a single signal. Other studies either analyze each signal separately, or fuse multiple signals into a composite index. This paper proposes a generic multisensor degradation model which can capture the correlation among sensors and learn the failure mechanism of units directly from multiple signals. As a result, the proposed method is more flexible and achieves better prognostic performance. 370F Vehicle Routing Sponsored: TSL, Freight Transportation & Logistics Sponsored Session Chair: Justin Goodson, Saint Louis University, St. Louis, MO, 63108, United States, goodson@slu.edu 1 - Generic Branch-and-Bound Schemes for Time Window Assignment Vehicle Routing Problems Anirudh Subramanyam, Carnegie Mellon University, DH3122, 5000 Forbes Ave., Pittsburgh, PA, 15213, United States, asubramanyam@cmu.edu, Chrysanthos Gounaris We study the allocation of long term delivery time windows to customers in the context of VRPs. Once a time window has been assigned, the operator must strive to meet it on a daily basis as well as possible. Since operational parameters such as customer demands or travel times vary from day to day, the aim is to assign time windows in a way that minimizes the expected routing cost. In this work, we show that this problem can be reduced to a variant of the Consistent VRP and adapt an algorithm we developed for the latter problem to solve to exact optimality instances of the former. Our approach is highly competitive and can address uncertainty in a wider class of parameters when compared to the previous state of the art. 2 - A Branch-Price-and-Cut Approach for Robust Vehicle Routing Akang Wang, Carnegie Mellon University, Pittsburgh, PA, 15213, United States, akangw@andrew.cmu.edu, Chrysanthos Gounaris We present a Branch-Price-and-Cut approach for tackling the Vehicle Routing Problem (VRP) under demand uncertainty. We utilize Limited-Memory Subset Row Cuts and robust Rounded Capacity Inequalities to strengthen the linear relaxation at every node. We use a Tabu Search heuristic to generate columns with negative reduced cost, and we resort to exact pricing via Dynamic Programming only if the former fails to identify such columns. Computational studies are performed on VRP benchmark instances stemming from multiple variants. 3 - Travel Time Correlation in Vehicle Routing Iurii Bakach, Graduate Student, University of Iowa, 2119 Keokuk Street, Apt 9, Iowa City, IA, 52240, United States, iurii-bakach@uiowa.edu, Ann Melissa Campbell, Jan Fabian Ehmke, Timothy Urban Due to the complexity of the problem, the majority of routing models consider arc costs to be deterministic and independent. However, in reality, arc costs are often correlated, for instance, when they reflect travel times in urban areas. We examine a routing problem with a min-max objective with both stochastic and correlated travel times. We develop a method using extreme-value theory to estimate the expected makespan and embed this within a routing heuristic. We present results that demonstrate the impact of different correlation patterns and levels of correlation on route planning. 4 - Route-Based Markov Decision Processes for Dynamic Vehicle Routing Problems Justin Goodson, Saint Louis University, John Cook School of Business, Davis-Shaughnessy Hall, St. Louis, MO, 63108, United States, goodson@slu.edu, Marlin Wolf Ulmer, Dirk C. Mattfeld, Barrett Thomas We propose a model for dynamic routing problems (DRPs) that extends the conventional Markov decision process (MDP) model for dynamic and stochastic optimization problems to more closely align with route-based solution methodologies in the DRP literature. We construct route-based MDPs by redefining the action space to operate on sets of planned routes. We generalize the state to include route plans and redefine the current-period reward/cost to be the marginal change in value associated with an update to a route plan. We show route-based MDPs are equivalent to the conventional MDP model. TE65

434