2015 Informs Annual Meeting

WE22

INFORMS Philadelphia – 2015

WE22 22-Franklin 12, Marriott Stochastic Processes II Contributed Session

3 - Efficient Monte Carlo Simulation using A Representation for Viscosity Solutions of Hamilton Jacobi Equations Pierre Nyquist, Brown University, 182 George Street, Providence, RI, 02912, United States of America, pierre_nyquist@brown.edu, Henrik Hult, Boualem Djehiche The design of efficient Monte Carlo methods is intimately connected to (sub)solutions of certain Hamilton-Jacobi (HJ) equations. However, such subsolutions can be hard to construct for a given problem. We discuss a min-max representation for viscosity solutions of HJ equations and its applications to rare- event simulation. To illustrate the result we focus on a specific setting: The use of importance sampling to estimate rare-event probabilities in a Markovian intensity model for credit risk. WE24 24-Room 401, Marriott Artificial Intelligence II Contributed Session Chair: Alexandra Diamond, Hunter College CUNY, 695 Park Avenue, New York, NY, 10021, United States of America, alexandradiamond@yahoo.com 1 - Role-assignment for Game-theoretic Cooperation In multiagent systems, a team of agents may work on multiple projects together. Here, we focus on the game-theoretic aspect that needs to be considered: when the agents are self-interested, careful role assignment is necessary to make cooperation the repeated game’s equilibrium. We formalize this problem and find an easy-to-check necessary and sufficient condition for a given role assignment to induce cooperation. We also show that finding whether such assignment exists is in general NP-hard. 2 - Discretization of Continuous Action Spaces in Extensive-form Games Christian Kroer, PhD Student, Carnegie Mellon University, 5000 Forbes Ave, Pittsburgh, PA, 15213, United States of America, ckroer@cs.cmu.edu, Tuomas Sandholm Most equilibrium-finding algorithms for sequential games require discrete, finite games, whereas applications often have continuity. Leveraging recent results on abstraction solution quality, we develop the first framework for providing bounds on solution quality for discretization of continuous action spaces in extensive- form games. Based on this framework, we develop algorithms for finding bound-minimizing abstractions under various structural assumptions. 3 - A New Fuzzy Kernel-free Support Vector Machine Model Jian Luo, Assistant Professor, Dongbei University of Finance and A new kernel-free fuzzy support vector machine model is proposed using the concept of Fisher discriminant analysis and a new fuzzy membership function. A decomposition algorithm is designed to solve this proposed model. Computational results indicate that the proposed model outperforms well-known fuzzy support vector machine models with kernels. 4 - Stochastic Approximation for Optimal Self-driven Sensor-guided Rescue Robots under Uncertainty Alexandra Diamond, Hunter College CUNY, 695 Park Avenue, New York, NY, 10021, United States of America, alexandradiamond@yahoo.com, Felisa Vazquez-abad A robot must locate and retrieve an artifact that has a limited battery for sending signals. We compare two scenarios. (1) All calculations are done in the cloud using signals from the artifact to estimate position. The robot is then sent to the rescue. (2) The robot has computational and sensing capabilities and it performs the optimization in real time as it moves towards the object. The “risk” is associated with the probability of missing the target by a certain amount. We discuss convergence Catherine Moon, Duke University, 213 Social Sciences, Box 90097, Durham, NC, 27705, United States of America, csm17@duke.edu, Vincent Conitzer Economics, 217 Jianshan Street, Shahekou District, Dalian, 116025, China, luojian546@hotmail.com, Zhibin Deng, Xingkai Yang, Huixiang Su, Chang Liu

Chair: Vincent Slaugh, Visiting Assistant Professor, Smeal College of Business, Penn State University, 465 Business Bldg, University Park, PA, 16802, United States of America, vws102@psu.edu 1 - Solving Strategic Refinery Planning and Financial Risk Problems via Pyomo Ariel Uribe, Ecopetrol S.A., Km 7 Via Piedecuesta, Piedecuesta, Colombia, ariel.uribe@ecopetrol.com.co, Wilson Briceño In this work we reformulate the problem proposed by Bagajewicz & Lakkhanawat (2008), in which is solved a refinery planning problem taking into account financial risk management issues and the pricing effect on the planning decisions. As a first step, we compare from a computational point of view the performance of the deterministic and stochastic solutions between gams and pyomo. Finally, we explore pyomo in a parallel environment. 2 - Valuation and Operation of Three Types of Power Plants using Continuous Time Stochastic Control Rune Ramsdal Ernstsen, University of Copenhagen, Finsensvej 42, 4.tv, Frederiksberg, 2000, Denmark, rre@math.ku.dk With the increasing focus on renewable energy in the deregulated energy markets, it is to be expected that the energy mix will change and along with it the dynamics of the energy prices. This will change the values of the existing and new power plants, and thus change the investment incentives. My research is based on valuation and operation of three different types of power plants in a new electricity market: a renewable power plant, a conventional power plant and a storage power plant. 3 - Reservation Admission Control in Rental Systems Vincent Slaugh, Visiting Assistant Professor, Smeal College of Business, Penn State University, 465 Business Bldg, University Park, PA, 16802, United States of America, vws102@psu.edu, Alan Scheller-wolf, Sridhar Tayur Using a multiserver queueing model, we study the problem of whether a rental firm should accept reservation requests during a rental season. We discuss performance bounds and the structure of the optimal policy for admitting reservations, and propose an easy-to-implement newsvendor-style heuristic for accepting reservations. We show that the heuristic performs well in test cases, and also observe that the system’s profit is decreasing in the reservation notice time. WE23 23-Franklin 13, Marriott New Directions in Applied Probability Sponsor: Applied Probability Sponsored Session Chair: David Goldberg, Assistant Professor, Georgia Institute of Technology, 755 Ferst Drive, NW, Atlanta, GA, 30332, United States of America, dgoldberg9@isye.gatech.edu 1 - Data-driven Dynamic Pricing and Inventory Control with Limited Price Changes Boxiao (beryl) Chen, University of Michigan-Ann Arbor, 1205 Beal Avenue, Ann Arbor, MI, 48109, United States of America, boxchen@umich.edu, Xiuli Chao We consider periodic review joint pricing and inventory control problem. Demand is random and price sensitive, and the firm has limited prior knowledge about its distribution. The firm learns the demand to maximize profit, but is faced with the business constraint that prevents it from conducting extensive price experimentations. We develop data-driven algorithms for pricing and inventory decisions and prove they converge to the optimal decision at the fastest possible speed. 2 - Rare Events in a Single Server Queue Harsha Honnappa, Perdue University, West Lafayette, IN, United States of America, honnappa@purdue.edu, Peter Glynn We characterize the rare event paths of the workload process in a FIFO single server queue that offers general service. The arrival times of a large, but finite, population of jobs are modeled as ordered statistics of i.i.d. random variables. We prove a large deviations principle (LDP) that shows that unlike a random walk, the most likely paths to hit a high level are not linear. We also extend this analysis to periodic traffic patterns, and develop an LDP for this setting as well.

486