Informs 2015

324 / 552

INFORMS Philadelphia – 2015
322
2 - Facility Location Problem with Appointment Scheduling
in Healthcare
Mengnan Chen, University of Central Florida, Department of
IEMS, Orlando, FL, 32816, United States of America,
, Qipeng Zheng
Facility location problem with appointment scheduling in healthcare is used to
schedule elective surgeries for physician, as well as to provide multiple choices for
patients. Facility location problem aims to improve the match between healthcare
resources (physician, clinic location) and patient needs (preferences and types of
diseases). By solving this problem, we can meet the minimum of loss for the
hospital (the total travel time) and satisfied the each patient preference as much
as possible.
3 - Exploiting Variability with Machine Learning Based Restart
Strategies for MIP Solvers
Lars Beckmann, University of Paderborn, Warburger Strafle 100,
Paderborn, De, 33100, Germany,
Performance variability, a.k.a. the heavy-tail phenomenon, can greatly affect the
solution time of combinatorial problems, especially mixed-integer programming
(MIP) models. We present a machine learning based restart strategy that exploits
variability for quickly solving MIP problems. We show that this approach is
effective on a large dataset of MIP models. Our approach can be integrated in any
MIP solver and can potentially be generalized for solving other combinatorial
problems as well.
4 - Faster Infeasibility Analysis for Mixed Integer Linear Programming
John Chinneck, Professor, Carleton University, Systems and
Computer Engineering, 1125 Colonel By Drive, Ottawa, ON,
K1S5B6, Canada, Andrew Scherr
Finding an Irreducible Infeasible Subset of Constraints in an infeasible mixed
integer linear program is extremely time consuming due to the need to solve
many different MILPs. We present a number of new algorithms that greatly speed
the process of analyzing infeasibility by reducing the number and size of MILP
subproblems solved.
5 - Algorithms for Determining Internal Credit Ratings
Srinivas Bollapragada, Chief Engineer, General Electric,
1 Research Circle, Niskayuna, NY, 12309, United States of
America, Xing Wang
Financial Institutions assign credit ratings to customers for managing risks.
Underwriters use these ratings to determine risk-based pricing. We developed an
efficient algorithm to assign credit ratings to customers. Our algorithm partitions
the probability of defaults associated with customers into a specified number of
risk classes to achieve a target mean probability of default for each class. GE
Capital Services uses our algorithm for its risk management needs.
TC13
13-Franklin 3, Marriott
Stochastic Combinatorial Optimization
Sponsor: Optimization/Optimization Under Uncertainty
Sponsored Session
Chair: Sean Skwerer, Yale, 300 George Street, Suite 523, New Haven,
CT, United States of America,
1 - Finite Horizon Markov Decision Problems and a Central Limit
Theorem for Total Reward
Alessandro Arlotto, Duke University, Durham, NC,
United States of America, , J. Michael Steele
We prove a CLT for a class of additive processes that arise naturally in the theory
of finite horizon Markov decision problems (MDPs). The theorem generalizes a
result of Dobrushin for temporally non-homogeneous Markov chains. The
innovation is that here the summands are permitted to depend on the current
state and a bounded number of future states of the chain. We show through
examples that this flexibility gives a direct path to asymptotic normality of the
total reward of finite horizon MDPs.
2 - Simplifying Ensembles of Trees
Sean Skwerer, Yale, 300 George Street, Suite 523,
New Haven, CT, United States of America,
Recursive partition functions (RPFs) are used in a wide variety of statistical
methods including classification and regression trees, regression splines, random
forests and boosting. The focus of this research is to find a framework for
aggregating an ensemble into a single tree or a small number of trees which have
comparable predictive strength to the entire ensemble. I will discuss advances
made in aggregating ensembles of recursive partition functions.
3 - Large Submatrix Detection in Gaussian Random Matrices
Quan Li, Massachusetts Institute of Technology, 550 Memorial
Drive Apt. 12B4, Cambridge, MA, 02139, United States of
America, , David Gamarnik
Iterative Search Algorithm is widely used to find large average submatrices of a
real-valued matrix in the exploratory analysis. It alternately updates rows and
columns until no further improvement is obtained. We present first theoretical
analysis of its performance in Gaussian random matrices. We show that the
algorithm terminates within finite iterations w.h.p.. This result implies that it
converges to a local maximum submatrix w.h.p., leading to a constant factor gap
from the global maximum.
4 - Locality in Optimization
Patrick Rbeschini, Yale University, New Haven, CT, United States
of America, Sekhar Tatikonda
How does the solution of a constrained network optimization problem behave
under perturbations of the constraints with respect to the topology of the
network? Typically, sensitivity results concern the objective function evaluated at
the optimal point, not the optimal point itself. We develop a general theory for
the local sensitivity of optimal points on networks. In the context of the network
flow problem, this theory yields that local perturbations on the constraints have
an impact on the components of the optimal point that decreases exponentially
with the graph-theoretical distance. These results suggest a notion of decay of
correlation for (non-random) optimization procedures. This notion can be used to
develop local algorithms that can substantially reduce the computational
complexity of canonical optimization procedures.
TC14
14-Franklin 4, Marriott
Stochastic Financial Optimization
Sponsor: Optimization/Optimization Under Uncertainty
Sponsored Session
Chair: Miguel Lejeune, Associate Professor, The George Washington
University, 2201 G Street, NW, Funger Hall, Suite 415, Washington, DC,
20052, United States of America,
1 - Scenario Decomposition for a Class of Risk-averse
Stochastic Programs
Pavlo Krokhmal, University of Iowa, 2136 Seamans Center,
Iowa City, IA, United States of America,
We consider nonlinear convex stochastic optimization problems where objective
or constraints involve downside coherent or convex risk measures of special form.
A scenario decomposition algorithm that exploits the constraint structure induced
by the corresponding risk measures is proposed. Numerical experiments on
portfolio optimization problems illustrate the computational effectiveness of the
developed procedure.
2 - Risk-budgeting Multi-portfolio Optimization with Portfolio and
Marginal Risk Constraints
Ran Ji, PhD Candidate, The George Washington University, 2201
G St, NW, Funger 415H, Washington, DC, 20052, United States of
America, Miguel Lejeune
We propose several stochastic risk budgeting multi-portfolio optimization models
with portfolio and marginal risk constraints. The models permit the simultaneous
optimization of multiple sub-portfolios with a downside risk measure assigned to
each asset and sub-portfolio. Each model includes a joint chance constraint with
random technology matrix. We expand a combinatorial modeling framework to
represent the feasible set of the chance constraint as a set of mixed-integer linear
inequalities.
3 - Factor-based Robust Indexing
Roy Kwon, Associate Professor, University of Toronto,
5 King’s College Road, Toronto, Canada,
Dexter Wu
We present an approach for tracking a benchmark index using robust factor
models. Robust versions of the Fama-French 3 and 5 factor models are developed
to construct uncertainty sets for a robust conic integer program. Constraints limit
risk, tracking error, and number of tickets. A Lagrangian-based strategy is
developed and computational results in tracking the SP 100 and SP 500 show that
robust indexing can offer enhanced indexation in turbulent market conditions.
TC13