2015 Informs Annual Meeting

TA24

INFORMS Philadelphia – 2015

TA22 22-Franklin 12, Marriott Two-Sided Matching Markets Sponsor: Applied Probability Sponsored Session Chair: Peng Shi, MIT Operations Research Center, 1 Amherst Street, E40-149, Cambridge, MA, 02139, United States of America, pengshi@mit.edu Co-Chair: Yash Kanoria, Assistant Professor, Columbia University, New York, United States of America, ykanoria@columbia.edu Co-Chair: Itai Ashlagi, MIT, 100 Main St, Cambridge, MA, 02139, United States of America, iashlagi@mit.edu 1 - On the Efficiency of Stable Matchings in Large Markets Sangmok Lee, Univ of Pennsylvania, 3718 Locust Walk, Philadelphia, PA, United States of America, sangmok@sas.upenn.edu, Leeat Yariv We study the wedge between stability and efficiency in large one-to-one matching markets. We show stable matchings are efficient asymptotically for a large class of preferences. In these environments, stability remains an appealing objective even on efficiency grounds, and monetary transfers are not necessary for efficiency purposes. Nonetheless, for severely imbalanced markets, when preferences entail sufficient idiosyncrasies, stable outcomes may be inefficient even asymptotically. 2 - Short Lists in Centralized Clearinghouses In the presence of frictions, participants in centralized clearinghouses generally fail to list all acceptable match partners. As a consequence, mutually acceptable pairs are left unmatched. The number of unmatched agents (and the happiness of matched agents) depends crucially on the structure of correlations in participants’ preferences. This work identifies a fundamental tradeoff between match quality and quantity, and uses this to offer guidance for the design of school choice mechanisms. 3 - How Much Choice is There in Two-sided Matching Markets? Itai Ashlagi, MIT, 100 Main St, Cambridge, MA, 02139, United States of America, iashlagi@mit.edu We study the structure of two-sided random matching markets with tiers. Our results provide insights on the amount of choice agents have in the core. TA23 23-Franklin 13, Marriott Asymptotic Optimality in Processing Networks Sponsor: Applied Probability Sponsored Session Chair: Itai Gurvich, Professor, Kellogg School of Management, Northwestern, 2001 Sheridan Rd., Evanston, IL, 60201, United States of America, i-gurvich@kellogg.northwestern.edu 1 - Approximations to Non-stationary Diffusion Processes Harsha Honnappa, Perdue University, West Lafayette, IN, United States of America, honnappa@purdue.edu, Peter Glynn Non-stationary diffusion processes emerge as limits to time inhomogeneous queueing processes in appropriately defined ‘high intensity’ regimes. In general, however, the transition densities of non-stationary diffusion processes are not known in closed form. Thus, in this talk, we present analytical approximations to expectations of these diffusion processes. This is joint work with Peter Glynn. 2 - On the Control of Fork-join Networks Erhun Ozkan, University of Southern California, Marshall School of Business, Los Angeles, CA, 90089, United States of America, eozkan@usc.edu, Amy Ward We study a prototypical fork-join network with two job classes and a shared server that processes both job types. We show that a cmu-type static priority policy is asymptotically optimal when the shared server is in some sense slow at processing the more expensive jobs. Otherwise, a state-dependent slow departure pacing control, under which the shared server sometimes gives priority to the less expensive jobs, is asymptotically optimal. 3 - Insensitivity and Optimality of Load Balancing with Processor Sharing Servers Varun Gupta, Varun.Gupta@chicagobooth.edu, Neil Walton We present some recent results and ongoing work on near-optimality and insensitivity properties of shortest queue load balancing under a carefully Nick Arnosti, Stanford University, Stanford, CA, United States of America, narnosti@stanford.edu

constructed many-servers asymptotic regime, when all the servers use the Processor Sharing scheduling rule. 4 - Capacity of Information Processing Systems Kuang Xu, Stanford University, United States of America, kuangxu@stanford.edu, Laurent Massoulie We study an information processing system where jobs are to be inspected by a set of experts. Inspections produce noisy results depending on the jobs’ hidden labels and the expert types, and an inspection occupies an expert for one time unit. The manager’s objective is to assign inspections so as to uncover the jobs’ hidden labels, using a minimum number of experts. Our main result is an asymptotically optimal inspection policy as the probability of error tends to zero.

TA24 24-Room 401, Marriott Intelligent Heuristics and Systems Sponsor: Artificial Intelligence Sponsored Session

Chair: Sam Thangiah, Professor/director, Slippery Rock University, Artificial Intelligence and Robotics Lab, 250 ATS, Slippery Rock, PA, 16057, United States of America, sam.thangiah@sru.edu 1 - A New Mathematical Model for Pattern Recognition in the Context of Feed Forward Neural Networks Sam Findler, Slippery Rock University, 510 Campus Side Cir, Slippery Rock, PA, 16057, United States of America, srf5132@gmail.com This paper sets out a new mathematical model for pattern recognition—in the form of what I call Pattern Recognition Circuits (PRCs) and Pattern Recognition Automata (PRAs). These new forms are given formal definitions and some preliminary theorems are proven. Next, the model is applied to feed forward neural networks, providing experimental grounding for the theory. Finally, the place of this new mathematical model in the context of general computational theory is discussed. 2 - Automatic Construction of Relational Features with Dataconda Michele Samorani, Assistant Professor, University of Alberta, 3-20F Business Building, University of Alberta, Edmonton, AB, T6G 2R6, Canada, samorani@ualberta.ca Traditional data mining and statistical techniques require a single table as input; by contrast, I tackle the problem of findings patterns in a set of related tables (Customers, Purchases, etc). This is made possible by Dataconda, a software freely available to academics which automatically generates a large number of features using information from all tables. In this talk, I will illustrate the benefits of this approach through an example in retailing. 3 - Massively Parallel GPU Accelerated Genetic Algorithm for Optimal Task Mapping in HPC Applications Ramanan Sankaran, Computational Scientist, Oak Ridge National Laboratory, P.O. Box 2008 MS 6008, Oak Ridge, TN, 37831, United States of America, sankaranr@ornl.gov Parallel applications on high performance computing (HPC) systems require network and topology aware task placement to ensure performance and scalability. We present a genetic algorithm for the quadratic assignment problem (QAP) that utilizes thousands of GPU accelerated compute nodes allocated for the application to compute an optimal task mapping in a few seconds. We show its convergence characteristics and impact on real life physics simulations on Titan, the most capable HPC system in the US. 4 - A Nao Humanoid Robot System for Interacting with Autistic Children Sam Thangiah, Professor/director, Slippery Rock University, Artificial Intelligence and Robotics Lab, 250 ATS, Slippery Rock, PA, 16057, United States of America, sam.thangiah@sru.edu, Daniel Martin, Michael Parnes, Mike Monfore, Stephen Fulton, Zachary Kearney, Brian Atwell, Justin Cather, Andrew Rindt, Sam Findler The NAO humanoid robot is a two feet tall robot with two hands and two feet with 25 degrees of freedom, cameras, microphones, sonar, infra-red and tactile and pressure sensors. It has various communication devices and an Intel ATOM processor. We describe a system implemented using the NAO robot to interact with autistic children. The system is designed to interact with autistic children in skill levels ranging from social to communication.

265