2016 INFORMS Annual Meeting Program

339 / 561

INFORMS Nashville – 2016
339
2 - Building Ensembles Of Support Vector Machine Classifiers
Through Multi Objective Optimization
Ozgu Turgut, Wayne State University, ,
Ratna Babu Chinnam
The trade-off parameter ‘C’ of support vector machine (SVM) classifiers is the key
input of training which tries to balance the generalization and empirical error.
However existing methods in determining this parameter usually based on brute
force search which is inefficient. The intend of this study is to incorporate a
genuine exact multi objective optimization algorithm that generates
representative Pareto optimal sets, with SVM in order to avoid C tuning phase.
Utilization of the algorithm for additional two major purposes is also studied;
namely, ensemble formation and confidence calculation of SVM classifiers.
3 - Multi-objective Optimization Under Multicollinearity
Haidar Almohri, Wayne State University, ,
Ratna Babu Chinnam, Arash Ali
While data driven process optimization methods are routine, several challenges
arise when dealing with variables with strong multicollinearity. In practice, there
might be multiple conflicting objectives as well, influenced by a common set of
variables. We propose optimization algorithms to jointly optimize multiple
objectives under multicollinearity. Methods are motivated by problems from
retailing and manufacturing domains.
TD17
105B-MCC
Statistical Learning with Convex Optimization
Sponsored: Optimization, Nonlinear Programming
Sponsored Session
Chair: Robert Freund, Massachusetts Institute of Technology, 100 Main
Street, Cambridge, MA, 02142-1347, United States,
Co-Chair: Rahul Mazumder, Massachusetts Institute of Technology, 100
Main Street, Cambridge, MA, 02142-1347, United States,
1 - An Optimal Aggregation Procedure For Nonparametric
Regression With Convex And Non-convex Models
Alexander Rakhlin, University of Pennsylvania,
Exact oracle inequalities for regression have been extensively studied in statistics
and learning theory over the past decade. In the case of a misspecified model, the
focus has been on either parametric or convex classes. We present a new
estimator that steps outside of the model in the non-convex case and reduces to
least squares in the convex case. To analyze the estimator for general non-
parametric classes, we prove a generalized Pythagorean theorem and study the
supremum of a negative-mean stochastic process (which we term the offset
Rademacher complexity) via the chaining technique. (joint work with T. Liang
and K. Sridharan)
2 - Convex Regularization For High-dimensional Tensor Regression
Ming Yuan, University of Wisconsin,
Data in the format of tensors, or multilinear arrays, arise naturally in many
modern applications. In this talk, I shall discuss several examples where convex
optimization approaches can be utilized for solving high-dimensional tensor
problems under low-dimensional structural assumptions.
3 - Distributed Proximal Algorithms For Convex Optimization
Garud Iyengar, Columbia University,
We propose a distributed first-order augmented Lagrangian algorithm to minimize
the sum of composite convex functions, where each term in the sum is only
known at one of the nodes, and only nodes connected by an edge can directly
communicate with each other. This optimization model abstracts a number of
applications in distributed sensing and machine learning. We show that any limit
point of the iterates is optimal; and for any epsilon>0, an epsilon-optimal solution
can be computed within O(log(1/epsilon)) iterations, whichrequire
O((d_max^{1.5}/d_min) /epsilon}) communications steps, where d_max (resp.
d_min) denotes the degree of largest (resp. smallest) degree node.
4 - Linear Estimation Through Unknown Non-linear Transformations
Constantine Caramanis, University of Texas,
Abstract to come.
TD18
106A-MCC
Resource Allocation Problems in Nonprofit Settings
Sponsored: Public Sector OR
Sponsored Session
Chair: Gemma Berenguer, Purdue University, 403 W State St, West
Lafayette, IN, 47907, United States,
1 - Public Facility Location And Fair Distribution Problem: Fractional
Programming Approach
Chong Hyun Park, Purdue University, West Lafayette, IN, 47906,
United States, , Gemma Berenguer
We consider a public facility location problem. A decision maker wants to
maximize the aggregated utilities of the service recipients while achieving the fair
distribution of distributed items. A bi-objective mixed integer linear fractional
programming problem is formulated and solved by various algorithms.
3 - Payment For Results: Funding Non-profit Operations
Sripad K Devalkar, Indian School of Business, Hyderabad, India,
Milind Sohoni, Neha Sharma
We consider the interaction between a donor making voluntary contributions to
fund a development project and the non-profit organization (NPO) using the
funds to implement the project. With information asymmetry about the NPO’s
efficiency and exogenous uncertainty affecting the actual benefit delivered by the
project, we show that ex-post funding (payment for results) may not always
maximize the donor’s utility even though it helps overcome information
asymmetry. We also characterize conditions under which ex-ante funding
(traditional funding) and ex-post funding (payment for results) are equilibrium
outcomes of the donor-NPO interaction.
4 - Using Optimization To Maximize Diversity In Engineering
Discussion Groups at The University Of Michigan
Kayse Lee Maass, University of Michigan, Ann Arbor, MI,
United States, , Mark Daskin
The College of Engineering at the University of Michigan assigns over 1250
freshmen to discussion groups for a “freshman reads” program. They want each
group to be as diverse as possible with respect to (a) country of origin, (b) gender,
(c) ethnicity, (d) in-state/out-of-state status, and (e) whether a student is the first
in his/her generation to attend college. We implemented a large-scale
optimization model to assign students to groups and have successfully used the
results for two years. The model and implementation are discussed.
TD19
106B-MCC
Stable Sets, Zero-forcing Sets, and Target Sets
in Graphs
Sponsored: Computing
Sponsored Session
Chair: Balabhaskar Balasundaram, Oklahoma State University,
322 Engineering North, Stillwater, OK, 74078, United States,
1 - Discrete vs. Continuous Formulations For Computing The
Stability Number: Results And Insights
Jitamitra Desai, Nanyang Technological University,
Jitamitra Desai, Nanyang Technological
University, Singapore, Singapore,
Balabhaskar Balasundaram
In this research, we compare and contrast discrete (0-1) and continuous
formulations for determining stable (independent) sets of a graph. In this context,
we define a new class of “vertex sets”, and derive an explicit characterization to
compute the number of alternate optima present in both the 0-1 and fractional
programming formulations. A byproduct of these vertex sets is another approach
to determine maximal independent sets. We also pose a conjecture that relates
the solution of the LP-relaxation to the 0-1 formulation to the number of
alternate optima to the stable set problem. Finally, some preliminary
computations on some standard testbed problems from the literature are
presented.
TD19