2015 Informs Annual Meeting

WD25

INFORMS Philadelphia – 2015

WD24 24-Room 401, Marriott Artificial Intelligence I Contributed Session Chair: Stanislaus Solomon, Assistant Professor Of Supply Chain Management, Sam Houston State University, 236S SHB, Department of Management and Marketing, Huntsville, TX, 77340, United States of America, solomon@shsu.edu 1 - Active Learning for Relevance Vector Machine Regression Youngdoo Son, PhD Candidate, Seoul National University, 1 Gwankak-ro, Gwanak-gu, Seoul, Korea, Republic of, hand02@snu.ac.kr, Jaewook Lee In this paper, we propose a novel active learning procedure for relevance vector machine regression. First, we propose the transductive relevance vector machine which uses both labeled and unlabeled data points to construct the model. Then, we suggest three simple querying strategies which exploit the characteristic of relevance vector machine regression. The proposed method showed significantly better performance than the benchmark, random selections, with both of artificial and real data sets. 2 - Rethinking Principal Component Analysis in EEG Classification Xiaoxia Li, North Dakota State University, 124 East Bison Court, Fargo, ND, 58102, United States of America, xiaoxia.li@ndsu.edu Principal Component Analysis (PCA) is considered to be a powerful tool in dimension reduction. However, it is worth thinking of the suitability of application for EEG signal data. Two EEG datasets collected from alcoholic and control groups were used to test the prediction accuracy before and after PCA transformation with SVM and KNN methods. Based on the classification results, we found that PCA is not valid in EEG signal processing. We also concern that other factors might be confounding. 3 - Regret Transfer and Parameter Optimization Noam Brown, Carnegie Mellon University, 5000 Forbes Ave, Pittsburgh, PA, 15213, United States of America, noamb@andrew.cmu.edu Regret matching is a widely-used algorithm for learning how to act. We begin by proving that regrets on actions in one game can be transferred and discounted to warm start the regrets for solving a different game with same structure but different payoffs that are a function of parameters. This provides, to our knowledge, the first principled warm-starting method for no-regret learning. We then leverage this warm starting to optimize a parameter vector for a player in a two-player zero-sum game. 4 - Dynamic Programming Approximations for Stochastic Resource Planning under Partial Observation Stanislaus Solomon, Assistant Professor Of Supply Chain Management, Sam Houston State University, 236S SHB, Department of Management and Marketing, Huntsville, TX, 77340, United States of America, solomon@shsu.edu, Cipriano Santos, Haitao Li, Keith Womer Assigning heterogeneous resources to jobs in a stochastic environment requires a sequential decision process. This research focuses on developing an approximate dynamic programming algorithm to address the problem scenario, where the information required to make decisions at each epoch is only partially observable. Hewlett-Packard’s resource planning problem will be studied as an application. WD25 25-Room 402, Marriott Joint Session AI/ICS: AI Planning and Operations Research Sponsor: Artificial Intelligence Sponsored Session Chair: Chris Beck, University of Toronto, 5 King’s College Rd, University of Toronto, Toronto, ON, M5S3G8, Canada, jcb@mie.utoronto.ca 1 - Heuristic Search in Dual Space for Constrained Stochastic Shortest Path Problems Sylvie Thiebaux, ANU & NICTA, 7, London Circuit, Canberra, Australia, Sylvie.Thiebaux@anu.edu.au, Felipe Trevizan, Pedro Santana, Brian Williams Constrained Stochastic Shortest Path Problems are a natural model for planning under uncertainty for resource-bounded agents with multiple competing objectives. We present i-dual, a heuristic search algorithm that generates optimal stochastic policies for these problems. It incrementally generates and explores promising regions of the ‘dual’ space of policy occupation measures, guided by a lower bound on the value function. Our experiments show significant run-time improvements over LP methods.

3 - A Storage Grid as a Network of “Internet of Things” Ehsan Shirazi, West Virginia University, Morgantown, WV, 26505, United States of America, ehshirazi@mix.wvu.edu This work focuses on a high-density storage system in which each grid works as an agent. Combining the decentralized, agent-based control with supervisory oversight provides a higher level of efficiency of product movement within the storage grid. Metrics in the research focus on variations from theoretical number of movements. 4 - Framework of Multiple Stochastic Decision Process in a System Toshikazu Aiyama, Professor, Tokyo Metropolitan University, 1-1 Minami Ohsawa, HachiOhji, Japan, tyaiyama@yahoo.com Consider multiple stochastic decision process operating in a system. A system is open to its environment; thus allowing external variability. We will concentrate on a two-process system on this research. First we will present various types. Next we will analyze some characteristics of each types. Some elementary numerical results are presented to illustrate the implication of more than one process in a system.

WD23 23-Franklin 13, Marriott Queueing Approximation and Simulation Sponsor: Applied Probability Sponsored Session

Chair: John Hasenbein, Mechanical Engineering, University of Texas at Austin, Austin, TX, United States of America, jhas@mail.utexas.edu 1 - Exact Simulation of Non-stationary Queues

Mohammad Mousavi, Assistant Professor, University of Pittsburgh, 1048 Benedum Hall, Pittsburgh, PA, 15212, United States of America, mousavi@pitt.edu

We discuss the challenges that arise in the planning simulations of systems with time dependent arrival and service rates. Estimating how far back in time a simulation must be initialized is an essential problem in planning simulations. We propound using reflected Brownian motion (RBM) with time-dependent drift and volatility as a guide for estimating this initialization time. We develop an exact simulation method for RBM with time-dependent drift and volatility. 2 - Heavy-traffic Limits for a Fork-join Network in the Halfin – Whitt Regime Hognyuan Lu, Penn State University, 355 Leonhard Bldg, University Park, PA, 16802, United States of America, hzl142@psu.edu, Guodong Pang We study a multi-server fork-join network with non-exchangeable synchronization (NES), where all parallel service stations are operating in the Halfin-Whitt regime under the non-idling FCFS discipline. The NES requires that completed tasks are only synchronized if they are associated with the same job. We prove FWLLN and FCLT for the number of tasks in each waiting buffer for synchronization, jointly with the number of tasks in each parallel service station and the number of synchronized jobs. 3 - Optimal Stock Allocation for Production-inventory Systems with Multiple Impatient Customer Classes Yasar Levent Kocaga, Assistant Professor Of Operations Management, Yeshiva University, 500 West 185th Street, New York, NY, 10033, United States of America, kocaga@yu.edu, Yen-Ming Lee We address the production and inventory control of a make-to-stock system with multiple impatient customer classes. We assume Poisson demand and exponential production times. Demand not satisfied immediately is backordered; but waits only up to an exponentially distributed amount of time, and is cancelled if not satisfied within this time. We show that the threshold inventory rationing policy is still optimal under certain conditions including a requirement on the order of abandonment rates. 4 - Routing and Scheduling in Fluid Gurvich Networks John Hasenbein, Mechanical Engineering, University of Texas at Austin, Austin, TX, United States of America, jhas@mail.utexas.edu, Arda Sisbot We examine a class of networks in which fluids may be routed to different classes at the same server. Extending previous work on single-station systems, we show that optimal policies in tandem Gurvich networks can exhibit interesting counter- intuitive behavior.

461