2016 INFORMS Annual Meeting Program

WD08

INFORMS Nashville – 2016

WD09 103B-MCC Spatial Optimization and Conservation Reserve Design Sponsored: Energy, Natural Res & the Environment I Environment & Sustainability Sponsored Session Chair: Bistra Dilkina, Georgia Institute of Technology, 266 Ferst Drive, Klaus Bldg 1304, Atlanta, GA, 30332-0765, United States, bdilkina@cc.gatech.edu 1 - Density Based Design: A Spatial Optimization Model For Protecting The Fisher Richard Church, University of California, Santa Barbara, church@geog.ucsb.edu We discuss the necessary elements of home range core areas in supporting female fishers during the key natal-maternal season. We propose an integer linear programming model that schedules harvests and spatially tracks and meets needed habitat elements over time and present preliminary results for an industrial forest in California. 2 - Optimizing Conservation Designs With Home Range And Connectivity Criteria Bistra Dilkina, Georgia Institute of Technology, Atlanta, GA, United States, bdilkina@cc.gatech.edu We develop a wildlife reserve design approach that takes into account both the number of individuals supported and the accessibility of the whole design to those individuals. In particular, a spatial capture-recapture model based on ecological resistance distance gives rise to three estimated quantities of interest (density, potential connectivity and density-weighted connectivity) that can be used to evaluate the ability of a reserve design to support individuals. We formulate a set of optimization problems to examine the use of these three quantities for selecting land parcels for conservation. 3 - Delaying Invasive Spread: Is Effective Control Possible Without Effective Prediction? Gwen Spencer, Smith College, gwenspencer@gmail.com Some (continuous) models of species spread yield impossibly-clean analytical results. Productive mathematical exchange with ecologists must acknowledge and attempt to capture disciplinary knowledge and critiques, even if this means sacrificing analytical traction. We will discuss computational work motivated by experimental and statistical papers in the invasive-species literature, making the case for discrete methods that acknowledge landscape heterogeneity and objectives that go beyond expected value. 4 - Avicaching: A Two Stage Game For Bias Reduction In Citizen Science Yexiang Xue, Cornell University, yexiang@cs.cornell.edu The data collected in citizen science projects are often biased, more aligned with the citizens’ preferences rather than scientific objectives. We introduce a novel game for reducing the data bias in which the organizer, a citizen-science program, incentivizes the agents, the citizen scientists, to visit under-sampled areas. We provide a novel way to encode this two-stage game as a single optimization problem, cleverly folding the agents’ problems into the organizer’s problem. We apply our methodology to eBird, a well-established citizen science program, as a game called Avicaching. Since its deployment, Avicaching has been very successful, surpassing the expectations of the eBird organizers. WD11 104A-MCC Various Aspects of Second Order Cone Optimization Sponsored: Optimization, Linear and Conic Optimization Sponsored Session Chair: Sertalp Bilal Cay, Lehigh University, 200 W Packer Ave, Bethlehem, PA, 18015, United States, sertalpbilal@gmail.com 1 - On Disjunctive Conic Cuts When They Exist When They Cut Mohammad Shahabsafa, Lehigh University, 200 West Packer Ave, Bethlehem, PA, Bethlehem, PA, 18015, United States, mos313@lehigh.edu, Tamas Terlaky The development of Disjunctive Conic Cuts (DCCs) for MISOCO problems has recently gained significant interest in the optimization community. Identification of cases when DCCs do not exist, or are not useful, saves computational time. In this study, we explore cases where either the DCC methodology does not derive a DCC which is cutting off the feasible region, or a DCC does not exist. Additionally, we work on extending the DCCs to other conic optimization problems such as Mixed Integer p-order Cone Optimization and Mixed Integer Semidefinite Optimization.

2 - Detecting Node Propensity Changes In Dynamic Degree- corrected Stochastic Block Model Lisha Yu, City University of Hong Kong, Hong Kong, lishayu2-c@my.cityu.edu.hk, Kwok-Leung Tsui Many real-world data can be represented as dynamic networks which are the evolutionary networks with timestamps. Studying the evolution of node propensity over time is significant to exploring and analyzing networks. In this paper, we propose a multivariate surveillance plan to monitor node propensity in dynamic social networks based on the degree-corrected stochastic block model (DCSBM). Experiments on simulated and real social network streams demonstrate that our surveillance strategies can efficiently detect different types of node propensity change in dynamic DCSBM with different kinds of community structure.

WD08 103A-MCC Dynamic Prog/Control Contributed Session

Chair: Xiaodong Luo, Sabre Holdings Inc, 3150 Sabre Drive, Southlake, TX, 76092, United States, Xiaodong.Luo@sabre.com 1 - Adoptive Vehicle Cruise Control Using Real Time Data And A Dynamic Programming Model: A Web Based Application Mohammad Ali Alamdar Yazdi, PhD Student, Auburn University, 354 W Glenn Ave, Auburn, AL, 36830, United States, mza0052@auburn.edu, Fadel Mounir Megahed Recent investigations have included exploring the benefits of autonomous vehicle systems in improving a vehicle’s miles per gallon (mpg) fuel-economy performance. This talk examines a new direction where large amounts of online data will be used to develop a fuel efficient cruise controller. More specifically, a dynamic programming model is constructed to capitalize on existing fuel consumption models and use real-time data collected from different Google APIs to minimize fuel efficiency. The minimization is based on dynamically optimizing the current speed and route based on forthcoming route conditions (traffic, elevation, etc.). 2 - Modeling Quality Of Care In Hospice Operations Leela Nageswaran, Tepper School of Business, 5000 Forbes Avenue, Pittsburgh, PA, 15213, United States, lnageswa@andrew.cmu.edu, Alan Scheller-Wolf, Aliza R Heching We study a hospice manager’s problem of controlling quality of care in light of recent regulatory changes mandating reporting of quality metrics. To explore the potential effects of such reporting, we develop an analytical Markov Decision Process model that incorporates how staffing - a primary determinant of quality - affects the rate at which patients join the hospice (due to quality reputation effects) and depart the hospice (due to quality of care effects). This in turn affects the hospice’s revenues and costs. We solve our model to obtain properties and insights related to the optimal quality control policy. 3 - Decision Facing Ambiguity MDP POMDP And Beyond Mohammad Rasouli, University of Michigan, 430 South Fourth Ave, Ann Arbor, MI, 48104, United States, rasouli@umich.edu While most of the decision making tools are developed for a Bayesian framework where the decision maker knows full stochastic description of uncertainties in the environment, decision facing ambiguity (model uncertainty and non-stochastic uncertainty) is a better approach for modeling a lot of practical situations. We discuss how decision making tools including MDP, POMDP, learning (e.g. Multi- armed bandit) and team decision making can be extended for environments with ambiguity. We discuss robustness and bounded rationality in this framework. 4 - Iterative Methods For Large Markov Decision Problems Xiaodong Luo, Sabre Holdings Inc, 3150 Sabre Drive, Southlake, TX, 76092, United States, Xiaodong.Luo@sabre.com We propose a new unified LP formulation for the Infinite Horizon Markov Decision Problem (MDP), both with the total discounted reward criteria and the long-run expected average reward criteria (assuming the gain is constant). We embed a column generation scheme into a multiplier method to solve the new formulation. Our algorithm can solve large randomly generated MDPs faster than commercial solvers. It scales up linearly for MDPs with hundreds of millions of nonzero. It uses much less memory than Barrier method, easy to warm start and is highly parallelizable.

458