Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

MB38

2 - Active Routing and Guidance Framework for Robust and Intelligent Taxiing Jun Chen, Queen Mary University of London, London, United Kingdom, Michal Weizser With increasing demand for air travel and overloaded airport facilities, inefficient airport taxiing operations are identified as a significant. An Active Routing and Guidance Framework is proposed for complex ground handling problems at major airports to reduce taxi times, operating costs and environmental impact. The problem is interdisciplinary and multi-dimensional as the efficiency of airport operations depends on aircraft dynamics, airport layout, air traffic, and other constraints. Results show that by considering all these factors in an integrated decision making framework the generated routing and guidance solutions are more realistic for pilots to follow, saving time and fuel. 3 - Dynamic Control of Airport Capacity Allocation with Stochastic Considerations Robert Shone, Lancaster University, Lancaster, United Kingdom, Kevin D. Glazebrook, Konstantinos G. Zografos The ever-increasing demand for air transport services continues to put the resources of the world’s busiest airports under tremendous pressure. Many optimization models have been proposed for runway scheduling, aircraft sequencing and other problems related to air traffic flow management. In this talk we discuss how to develop new approaches which explicitly take into account uncertainty caused by weather conditions and other operational factors. n MB37 North Bldg 225A Machine Learning and Causality Sponsored: Applied Probability Sponsored Session Chair: Nathan Kallus, Cornell University, New York, NY, 10044, United States Co-Chair: Stefan Wager, Stanford GSB, Stanford GSB, New York, NY, 10023-2154, United States 1 - Interpreting Predictive Models for Human-in-the-loop Analytics Hamsa Sridhar Bastani, Wharton School, Philadelphia, PA, United States, Osbert Bastani, Carolyn Kim Interpretability has become an important issue as machine learning is increasingly used to inform consequential decisions. We propose an approach for interpreting complex blackbox models by extracting a decision tree that approximates the model. The algorithm avoids overfitting by actively sampling new training points using the blackbox model. We use this technique to interpret a random forest classifier for predicting diabetes risk. Physicians successfully used our interpretation to discover an unexpected causal issue in the diabetes classifier. 2 - Long Tail Phenomenon in Discrete Choice Estimation Pu He, Columbia University, Uris Hall, Cub 4H, New York, NY, 10027, United States, Fanyin Zheng Long tail distribution of sales or market share data is a common phenomenon in empirical studies in economics, operations, and marketing. Classic discrete choice estimation framework ignores the long tail and can lead to biased estimates. In this work, we introduce a new two-step procedure to solve the problem. Our solution applies machine learning algorithms to estimate market shares in the first stage, and in the second stage we estimate a weighted multinomial logit model to recover customer preference parameters. We show that our proposed approach corrects for the bias in demand estimates and improves profits when these estimates are used in pricing decisions. 3 - Learning to Personalize Safely Under Unobserved Confounding Nathan Kallus, Cornell University, 2 W. Loop Road, # 316, New York, NY, 10044, United States Recent work on counterfactual learning from observational data aims to leverage large-scale data — much larger than any experiment can ever be — to learn individual-level causal effects for personalized interventions. The hope is to transform electronic medical records to personalized treatment regimes, transactional records to personalized pricing strategies, and click- and “like”- streams to personalized advertising campaigns. Motivated by the richness of the data, existing approaches make the simplifying assumption that there are no unobserved confounders: unobserved variables that affect both treatment and outcome and would induce non-causal correlations that cannot be accounted for. However, all observational data, which lacks experimental manipulation, no matter how rich, will inevitably be subject to some level of unobserved confounding and assuming otherwise can lead to personalized treatment policies that seek to exploit individual-level effects that are not really there, may intervene where not necessary, and may in fact lead to net harm rather than net good relative to current, non-personalized practices.

4 - Quasi-oracle Estimation of Heterogeneous Treatment Effects Stefan Wager, Stanford University, 655 Knight Way, Stanford, CA, United States, Xinkun Nie Flexible estimation of treatment effects lies at the heart of many statistical challenges. We develop a class of algorithms for heterogeneous treatment effect estimation in observational studies. We estimate marginal effects and treatment propensities to form an objective function that isolates the causal signal. Then, we optimize this objective function. For both steps, we can use any loss-minimization method fine-tuned by cross validation. In the case of penalized kernel regression, we show that even if the pilot estimates for marginal effects and treatment propensities are not particularly accurate, we achieve the same regret bounds as an oracle who has a priori knowledge of them. n MB38 North Bldg 225B Joint Session APS/ENRE: Stochastic Modeling and Optimization for Energy Systems Sponsored: Applied Probability Sponsored Session Chair: Adam Wierman, California Institute of Technology, Pasadena, CA, 91125, United States Co-Chair: Alessandro Zocca, California Institute of Technology, Pasadena, CA, 91125, United States 1 - Power Grid State Estimation Following Cyber-physical Attacks Saleh Soltan, Princeton University, NJ, United States, Mihalis Yannakakis, Gil Zussman, Prateek Mittal, H. Vincent Poor I will provide a summary of our recent results on power grid state estimation following cyber-physical attacks in which an adversary attacks an area by: (i) disconnecting some lines within the attacked area, and (ii) blocking/modifying the measurements from monitoring devices within the area to mask the line failures. We demonstrate that by using tools from graph theory and by leveraging the algebraic properties of the power flow equations, one can detect the attacked area as well as line failures in polynomial time under some topological constraints on the area. I then show that stochastic version of these methods can be used to detect line failures efficiently in general attacked area topologies. 2 - Understanding the Inefficiency of Security-constrained Economic Dispatch Enrique Mallada, Johns Hopkins University, 3400 N. Charles St, Barton Hall 312, Baltimore, MD, 21218, United States, Mohammad Hajiesmaili, Wuhan Desmond Cai The security-constrained economic dispatch (SCED) problem tries to maintain the reliability of a power network by ensuring that a single failure does not lead to a global outage. In this talk, we analyze the economic cost of incorporating security constraints in economic dispatch. Inspired by inefficiency metrics in game theory, we introduce the notion of price of security as a metric that formally characterizes the economic inefficiency of SCED. We investigate the impact of generation availability and demand distribution on the price of security. Our results show that renewable sources, with nearly zero marginal costs, can have a high impact on the price of security. 3 - Robust Volt-var Optimization in Power Distribution Systems Weixuan Lin, Cornell University, Ithaca, NY, United States, Eilyan Bitar We consider the decentralized reactive power control of photovoltaic (PV) inverters spread throughout a radial distribution network. Our objective is to minimize the expected voltage regulation error, while guaranteeing the robust satisfaction of distribution system voltage magnitude and PV inverter capacity constraints in real-time. We provide a method to compute a robust decentralized controller via the solution of a finite-dimensional conic program. The resulting trajectories of PV inverter reactive power injections and nodal voltage magnitudes are guaranteed to be feasible for any realization of the system disturbance under the proposed control policy. 4 - Failure Localization in Power Systems via Tree Partitions Alessandro Zocca, California Institute of Technology, 1200 E. California Blv, MC 305-16, Pasadena, CA, 91125, United States, Linqi Guo, Chen Liang, Steven Low, Adam Wierman Cascading failures in power systems propagate non-locally, making the control and mitigation of outages extremely hard. Using the emerging concept of the tree partition, we provide a complete analytical characterization of line failure localizability in transmission networks. The crucial insight is that power systems that have more and thus smaller tree partition components are less vulnerable to large-scale outages. Furthermore, our characterization suggests that by switching off only a negligible portion of transmission lines we can have a significantly better control of cascading failures without significantly increase line congestion across the network.

161