Informs 2015

74 / 552

INFORMS Philadelphia – 2015
72
SB20
20-Franklin 10, Marriott
Cloud Resource Management and Pricing
Cluster: Cloud Computing
Invited Session
Chair: Daniel Grosu, Wayne State University, 5057 Woodward Ave,
Detroit, United States of America,
Co-Chair: Lena Mashayekhy,
1 - Efficient and Online Reconfigurations in NoSQL Databases
Indranil Gupta, Associate Professor, University of Illinois at
Urbana-Champaign, 201 N. Goodwin Ave., Urbana, IL, 61801,
United States of America, Mainak Ghosh,
Yosub Shin, Wenting Wang, Gopalakrishna Holla
NoSQL databases are soon to be a multi-billion dollar market. In live NoSQL
deployments “reconfiguration operations,” affecting a lot of data at once, are a
major pain point, e.g., changing a table’s primary key. We describe techniques to
perform reconfiguration in NoSQL systems (sharded and ring-based) quickly and
online, i.e., while supporting high availability and low latency for reads/writes.
We present results from our implementation and deployment with popular open-
source NoSQL systems.
2 - Exploring Market Models for Software-defined Systems
Manish Parashar, Rutgers University, 110 Frelinghuysen Road,
Piscataway, NJ, 08854, United States of America,
Software-defined platforms, such as those enabled by Cloud services, provide new
levels of flexibility, which can lead to very dynamic infrastructures that adapt
themselves to application needs. In this talk I will explore how a cloud-of-clouds
marketplace can support data-driven applications by programmatically federating
geographically distributed resources to satisfy QoS demands. Specifically, I will
focus on market and utility based models that can drive the synthesis of such
systems.
3 - Stochastic Optimal Control of Time-varying Workloads
Mark Squillante, IBM Research, Thomas J. Watson Research
Center, 1101 Kitchawan Road, Yorktown Heights, NY,
United States of America, Yingdong Lu,
Mayank Sharma, Bo Zhang
Motivated by cloud computing, we consider GI/GI/1 queueing systems under
time-varying workloads on one time scale and under time-varying controls on
another time scale. We derive structural properties for the optimal dynamic
control policy in general, establishing that this policy can be obtained through a
sequence of convex programs. We also derive fluid and diffusion approximations
and solutions for the problem. Computational experiments demonstrate the
benefits of our approach.
SB21
21-Franklin 11, Marriott
Healthcare Data Analytics
Sponsor: Health Applications
Sponsored Session
Chair: Donald Lee, Yale School of Management, 165 Whitney Ave,
New Haven, CT, United States of America,
1 - A Statistical Approach to Cost-effectiveness Analysis under
Uncertainty about the Willingness-to-pay
Reza Yaesoubi, Yale School of Public Health, 60 College Street,
New Haven, CT, 06510, United States of America,
, Forrest Crawford, David Paltiel
Although it plays a central role in cost-effectiveness analysis, societies’ willingness
to invest for an additional unit of health is rarely known to policy makers. In this
work, we develop a statistical model to help decision makers determine whether a
new healthcare alternative is considered cost-effective in the absence of exact
value for the willingness-to-pay for health.
2 - Networks Classification via Mathematical Programming
Daehan Won, University of Washington, Seattle, 1415 NE
Ravenna Blvd, #401, Seattle, WA, 98105, United States of
America,
We are developing mathematical programming models to classify the network
structured data. Along the line with the feature selection, we present node
selection approach to increase the classification accuracy as well as improving
interpretability. To verify the utility of our proposed approach, we demonstrate
the result of brain functional connectivity network data set.
3 - Outcome-driven Personalized Treatment Design for
Managing Diabetes
Eva Lee, Georgia Institute of Technology,
This work is joint with Grady Memorial Hospital and the Atlanta VA Medical
Center. We discuss an evidence-based decision support tool that couples a
treatment predictive model with a planning model. Specifically, the predictive
model uncovers drug effect based on pharmaco-kinetics and dynamics analysis.
This evidence is then modeled within the personalized planning model for
optimal treatment plan design. Results for a collection of patients will be
presented.
SB22
22-Franklin 12, Marriott
Experiment Design and A-B Testing
Sponsor: Applied Probability
Sponsored Session
Chair: Ciamac Moallemi, Columbia Business School, 3022 Broadway,
Uris Hall, New York, United States of America,
Co-Chair: Vivek Farias, Associate Professor, MIT, 100 Main Street,
Cambridge, MA, United States of America,
1 - Online A-B Testing
Vivek Farias, Associate Professor, MIT, 100 Main Street,
Cambridge, MA, United States of America,
Nikhil Bhat, Ciamac Moallemi
We consider the problem of optimal A-B testing when the impact of a treatment
is marred by a large number of covariates and subjects arrive and are assigned to a
treatment sequentially. Our objective is to maximize the efficiency of our estimate
of the treatment effect. Our main contribution is to show that what is typically
thought of as a high-dimensional, intractable problem is, in fact, tractable under
reasonable assumptions.
2 - Covariate Balanced Restricted Randomization: Optimal Designs,
Exact Tests, and Asymptotic Results
José Zubizarreta, Columbia Business School, 3022 Broadway, Uris
Hall, Room 417, New York, NY, 10027, United States of America,
, Jingjing Zou
We present a new method for the design of randomized experiments that (i) uses
integer programming to optimally match subjects before randomization, (ii)
assigns matched subjects to treatment and control in a randomized but controlled
or restricted fashion, (iii) yields an explicit exact randomization distribution for
small samples, and (iv) has good asymptotic properties in large experiments. We
illustrate this method with a real experiment and show its advantages beyond
standard methods.
3 - Can I Take a Peek? Continuous Monitoring of A/B Tests
Ramesh Johari, Stanford University, 475 Via Ortega, Stanford, Ca,
94305, United States of America, ,
Leo Pekelis, David Walsh
Statistical results for A/B tests are computed under the assumption that the
experimenter will not continuously monitor their test. If users continuously
monitor experiments, as is common in practice, then what statistical methodology
is appropriate for hypothesis testing, significance, and confidence intervals? We
present recent work addressing this question, building from results in sequential
hypothesis testing. This work was carried out with Optimizely, a leading A/B
testing platform.
SB20