Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TB08

4 - Near-Optimal Joint Matching via Convex Relaxation Yuxin Chen, Princeton University, NJ, United States

5 - A Stochastic Optimization Model for Road Network Protectionand Restoration During a Natural Disaster Sachin Mhatre, Graduate Assistant, North Carolina Agricultural & Technical State University, 1601, E. Market St., Greensboro, NC, 27411, United States, Xiuli Qu Natural disasters disrupt infrastructures, communities and thousands of individuals every year in the world. Effective protection and restoration of road transportation systems play an important role in disaster operations, such as rescuing and relief delivery. In this study, a stochastic integer linear programming model is developed to schedule road protection activities and locate road restoration resources when a hurricane approaches, and to plan road restoration activities and allocate resources after a hurricane. The model integrates the decisions for both preparedness and response activities during a disaster, which will improve decision making in disaster management. n TB08 North Bldg 124A Optimization for Robust and Risk-Aware Learning Sponsored: Optimization/Nonlinear Programming Sponsored Session Chair: Alec Koppel, U.S. Army Research Laboratory,Silver Spring, MD, 20910, United States 1 - Models and Algorithms for Risk-sensitive Inverse Reinforcement Learning Marco Pavone, Phd, Stanford University, Palo Alto, CA, 94305, United States The literature on Inverse Reinforcement Learning (IRL) typically assumes that humans take actions in order to minimize the expected value of a cost function, i.e., that humans are risk neutral. Yet, in practice, humans are often far from being risk neutral. In this talk, I will present a framework for risk-sensitive IRL that can explicitly account for a human’s risk sensitivity. Specifically, I will present non-parametric and semi-parametric algorithms for inferring a human’s underlying risk measure and cost function for a rich class of static and dynamic decision-making settings. 2 - Risk-averse Multi-armed Bandit Problems Qing Zhao, Professor, Cornell University, 325 Rhodes Hall, Cornell University, Ithaca, NY, 14853, United States, Sattar Vakili Risk-averse multi-armed bandit problems are studied under the measure of mean-variance. Compared with the risk-neutral objective of maximizing the expected cumulative reward, the mean-variance measure defined on the observed reward process leads to a regret that depends on high-order statistics of the random time spent on each arm and can no longer be written as a simple sum of certain immediate performance loss defined in isolation for each time instant. Regret lower bounds and order-optimal policies are established within both problem-specific and minimax frameworks. 3 - Controlling the Bias-Variance Tradeoff via Coherent Risk for Robust Online Learning with Kernels Alec Koppel, Phd, U.S. Army Research Laboratory, Adelphi, MD, 20783, United States In supervised learning, we learn a statistical model by minimizing a merit of fitness averaged over data and ignore the variance, i.e., the gap between the optimal within a hypothesized function class and Bayes Risk. We account for *both* by incorporating coherent risk, which quantifies decision uncertainty. We develop the first solution to this problem when in reproducing kernel Hilbert spaces (RKHS) called Compositional Online Learning with Kernels (COLK). COLK uses stochastic quasi-gradient together with greedy projections to mitigate the per-iteration complexity of RKHSs. We establish COLK converges and that its complexity is at-worst finite. Experimentally, COLK overcomes overfitting. 4 - Recursive Optimization of Convex Risk Measures: Mean-Semideviation Models Dionysios Kalogerias, Postdoctoral Research Associate, Princeton We develop and analyze stochastic subgradient methods for optimizing a new class of convex risk measures, termed as Mean-Semideviations. First, we study the basic properties of Mean-Semideviations, and constructively characterize them from a theoretical viewpoint. We then introduce the MESSAGEp algorithm, a compositional stochastic subgradient procedure for solving convex Mean- Semideviation risk-averse problems to optimality. Pathwise convergence of the MESSAGEp algorithm is established within a new, versatile and powerful theoretical framework, which reveals a fundamental tradeoff between the smoothness of the cost function and that of the particular risk measure of choice. University, Sherrerd Hall Room 119, 98 Charlton Street, Princeton, NJ, 08544, United States, Warren B. Powell

Joint matching over a collection of objects aims at aggregating information from a large collection of similar instances (e.g. images, graphs, shapes) to improve maps between pairs of them. Given multiple matches computed between a few object pairs in isolation, the goal is to recover an entire collection of maps that are (1) globally consistent, and (2) close to the provided maps – and under certain conditions provably the ground-truth maps. In this work, we develop a convex relaxation algorithm to jointly match multiple objects that exhibit only partial similarities, given a few pairwise matches that are densely corrupted. The algorithm provably exhibits near-optimal error-correction ability. n TB07 North Bldg 123 Networked Infrastructure Resiliency and Recovery Sponsored: Optimization/Network Optimization Sponsored Session Chair: Sachin Mhatre, North Carolina A&T State University, North Carolina A&T State University, Greensboro, NC, 27408, United States 1 - System Resilience Modeling Under Multi-hazard Yao Cheng, Beihang University, 37 Xueyuan Road, Haidian District, Beijing, 100191, China There has been significant development of complex engineered systems in the last two decades. The increasing natural and manmade hazards have caused significant performance disruption of such systems. In this presentation, we focus on one of the extended reliability metrics, namely resilience to assess the ability of such systems to withstand and recover when hazards occur. We propose quantifications of resilience for non-repairable and repairable systems under multi hazard. As restoration is conducted with limited resource under most circumstances, we recommend importance measures (IMs) to identify critical components in the system. 2 - Resilience Investment in Supply Chain Network Under Stochastic Disruptions Kedong Chen, University of Minnesota, Minneapolis, MN, 55454, United States,, Ankur Mani, Kevin W. Linderman Supply chain disruptions are common yet costly. Companies are motivated to invest resilience resource in their supply chain networks to mitigate disruptions, but they face the question of which nodes to invest. This study characterizes the optimal strategies of resilience investment for generic layered supply chain networks. Through analyzing the stochastic network where resilience investment reduces the chance of a node being disrupted, we find node capacity and path characteristics (path length and flow) to be key factors in the investment decision. The frequency of disruptions determines the management focus between node and path. Managerial implications are provided. 3 - Bi-objective Restoration Plan Optimization for Transportation Infrastructures Considering Both Total Travel Time and Unmet Demand Tingting Zhao, University of South Florida, 4202 E. Fowler Avenue, ENB 118, Tampa, FL, 33612, United States A bi-objective bi-level problem is formulated for restoration plan optimization of transportation infrastructure system after disruptions. The upper-level one is to maximize system resilience measured by total travel time and unmet demand with limited resources. The lower-level one is a network flow assignment problem considering unmet demand in the system after the event. The single objective problem can be linearized and then transformed to MILP problem. The bi- objective MILP problem is solved by Triangle Splitting Method. The proposed methodology is evaluated with a typical road network. 4 - A Heuristic Approach to the Post-disturbance Microgrid Formation Problem Xin Shi, Lehigh University, Bethlehem, PA, United States, Kwami Senam A. Sedzro, Alberto J. Lamadrid, Luis F. Zuluaga Microgrid formation is a potential solution in post-disaster electric grid recovery efforts. Recent works propose distribution level microgrid formation models using MILP techniques. However, these models can only be solved for small and medium-size power systems due to their computational intractability. In this talk, we introduce a heuristic approach that allows to approximately solve the microgrid formation problem for medium to large, more realistic, instances. Furthermore, the proposed approach allows to approximately solve the stochastic version of the problem.

277