Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

MC37

4 - A New Strategic Airline Fleeting and Pricing Optimization Model Xiaodong Luo, Sabre Holdings Inc, 3150 Sabre Drive, Southlake, TX, 76092, United States We study a new integrated airline fleet assignment model when itinerary prices are also subject to change. Both fleet assignment models with complex passenger mix behavior as well pricing optimization models are well studied in the literature, but they are rarely studied simultaneously. We will use a combination of nonlinear optimization, heuristics and integer programming techniques to reduce computational time while still obtain close to optimal solutions. We will use discrete event simulation based on real airline data to demonstrate the power of our approach. n MC37 North Bldg 225A Joint Session APS/ENRE: Applied Probability and Power Systems Sponsored: Applied Probability Sponsored Session Chair: Bert Zwart, Eindhoven, 5629RD, Netherlands Co-Chair: Foekje Sloothaak, Eindhoven University of Technology, Eindhoven, 5612 AZ, Netherlands 1 - Battery Swapping Stations for Electric Vehicles: A Queueing Perspective Foekje Sloothaak, Eindhoven University of Technology, Den Dolech 2, Eindhoven, 5612 AZ, Netherlandsl Although there has been an increasing penetration of EVs in the last decade, the adoption of this technology remains slow, partly due to issues with long battery charging times. We consider the concept of battery swapping stations where EV users can quickly exchange their (almost) depleted batteries by full batteries. We take a queueing perspective by modeling this framework as a closed system operating under the QED policy. This policy yields favorable effects: EV users experience low waiting times, while battery swapping stations do not needlessly keep many spare batteries. Moreover, we show state-space collapse when EV users are inclined to swap their batteries at the station that is least loaded. 2 - Stochastic Networks for Electric Vehicle Charging Angelos Aveklouris, Eindhoven University of Technology, P.O. Box 513, Eindhoven, 5600 MB, Netherlands, Maria Vlasiou, Bert Zwart We analyze a stochastic system that models the performance of electric vehicle charging. The model takes into account the stochastic behavior of electric vehicles and the physical constraints in a low-voltage distribution grid. We model this as a class of resource-sharing networks and characterize the performance of the system by a fluid approximation. 3 - Chance-constrained AC Optimal Power Flow Line Roald, Assistant Professor, University of Wisconsin-Madison, 1645D 16th Street, Los Alamos, NM, 87544, United States Renewable electricity generation increase uncertainty in power system operation and necessitates new methods for planning and operation. We adopt a chance- constrained AC optimal power flow formulation, which guarantees constraint satisfaction with a pre-defined probability. Obtaining solutions for this problem is challenging due to the AC power flow equations, a set of non-linear, non-convex equality constraints that must be satisfied with high probability. We discuss two different solution approaches based on partial linearization and polynomial chaos expansion, discuss their respective drawbacks and advantages, and show numerical results for different IEEE test cases. n MC38 North Bldg 225B APS Session Title VII Sponsored: Applied Probability Sponsored Session Chair: Bo Zhang Co-Chair: Gauri Joshi, Carnegie Mellon University 1 - Fast Distributed Machine Learning in the Presence of Slow and Stale Updates Gauri Joshi, Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, PA, 15213, United States Stochastic Gradient Descent when distributed across multiple nodes, suffers from delays due to slow or straggling nodes. Asynchronous methods can alleviate stragglers, but cause gradient staleness that can adversely affect convergence. We present a theoretical characterization of the speed-up offered by asynchronous methods by analyzing the trade-off between the error in the trained model and the wall-clock training runtime. Our runtime analysis considers random straggler

delays, which helps us design and compare distributed SGD algorithms that strike a balance between stragglers and staleness. We also present a novel learning rate schedule to compensate for gradient staleness. 2 - Asynchronous Algorithms Faster Iterations Same Quality Robert Hannah, UCLA, Wotao Yin In this talk, we discuss recent results on the performance of asynchronous optimization algorithms. Using renewal theory we analyze how subproblem and computing power heterogeneity, random delays, and other factors lead to asynchronous algorithms completing faster iterations. In particular, under a standard model, random delays lead to a slowdown by a factor of ln(p), when the number of computing nodes p is sufficiently large. We then present results that show that a number of asynchronous algorithms need essentially the same number of iterations as their traditional counterparts to converge to a solution. And hence we conclude asynchronous algorithms will domination traditional ones at scale. 3 - Onthe Loss Surface of Neural Networks for Binary Classification R Srikant, University of Illinois, Coord Science Lab & Dept of ECE, University of Illinois, Urbana, IL, 61801, United States Deep neural networks used for classification problems are trained so that their parameters achieve a local minimum of an appropriate loss function. The loss function is typically intended to approximate the classification error in training samples. In this talk, we will show that approximations of widely-used neural network architectures have the property that every local minimum of a surrogate loss function is a global minimum, and further achieves the global minimum of the training error. Joint work with Shiyu Liang, Ruoyu Sun, and Jason Lee. n MC39 North Bldg 226A Networks, Reliability & Extremes Sponsored: Applied Probability Sponsored Session Chair: Pieere L’Ecuyer, Universite de Montreal, C.P. 6128, Succ Centre- Ville, Montreal, QC, H3C 3J7, Canada Co-Chair: Guido Lagos, Universidad de Chile, Santiago, 8370213, Chile 1 - A Concentration Phenomenon for Network Reliability under Dependent Failures Guido Lagos, Universidad de Chile, Jose Miguel Carrera 439, depto 802, Santiago, 8370213, Chile Guido Lagos, Universidad Adolfo Ibanez, Santiago, Chile, Javiera Barrera We consider the reliability of a network where link failures are correlated, where we define the reliability at a given time instant as the probability of, at that time, there being at least k fixed links working. Our main contribution is a concentration result of a non-trivial scaling regime for the reliability of the network, as the time and size of network scales. Our results allow to study the common-cause failure models on networks in a realistic, relevant, yet practical, fashion: it allows to capture correlated components in the network; it allows to estimate and give error bounds for the failure probabilities of the system; and at same time only needs to specify a reduced family of parameters. 2 - Efficient Monte Carlo Estimation of Network Reliability Metrics with the Standard Estimator Gerardo Rubino, INRIA, Campus de Beaulieu, Rennes, 35042, France We consider the network reliability problem, in a static context. Estimating the system reliability using the standard or crude approach is a trivial task. It becomes a hard problem when dealing with a rare event situation (that is, when the system reliability is close to one), because the standard approach is quickly useless. In this presentation we will describe a way of using the standard estimator in an efficient way, in the case of rare events. One interest of this idea is that it allows, in a direct way, estimating all kinds of metrics without changing the method, and also to estimate sensitivities. The talk will be illustrated with some realistic examples of networks. 3 - Large Deviations of Multivariate Gaussian Extrema Harsha Honnappa, Purdue University, 315 N. Grant Street, West Lafayette, IN, 47906, United States, Raghu Pasupathy, Prateek Jaiswal We present a large deviations (LD) analysis of multivariate Gaussian extreme value random vectors. Our analysis contrasts with classic extreme value theory, which focuses on the ‘nominal’ behavior of extreme values. We demonstrate that the LD behavior of the multivariate Gaussian extremes is Frechet-like. Furthermore, we identify the “dominating point” at which the extreme values enter convex regions of interest in the range of the extreme value random vectors. Finally, we discuss applications of the LD results to reliability problems.

198