Topics at the interface of optimization and statistics
Embargo until
Date
2020-07-22
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Johns Hopkins University
Abstract
Optimization has been an important tool in statistics for a long time. For example, the problem of parameter estimation in a statistical model, either by maximizing a likelihood function or using least squares approach, reduces to solving an optimization problem. Not only has optimization been utilized in solving traditional statistical problems, it also plays a crucial role in more recent areas such as statistical learning. In particular, in most statistical learning models, one learns the best parameters for the model through minimizing some cost function under certain constraints.
In the past decade or so, there has been an increasing trend in going to reverse direction: Using statistics as a powerful tool in optimization. As learning algorithms become more efficient, researchers have focused on finding ways to apply learning models to improve the performance of existing optimization algorithms. Following their footsteps, in this thesis, we study a recent algorithm for generating cutting planes in mixed integer linear programming problems and show how one can apply learning algorithms to improve the algorithm.
In addition, we use the decision theory framework to evaluate whether the solution given by the sample average approximation, a commonly used method to solve stochastic programming problems, is ``good". In particular, we show that the sample average solution is admissible for an uncertain linear objective over a fixed compact set and for a convex quadratic function with an uncertain linear term over box constraints when the dimension is less than 4.
Finally, we combine tools from mixed integer programming and Bayesian statistics to solve the catalog matching problem in astronomy, which tries to associate an object's detections coming from independent catalogs. This problem has been studied by many researchers. However, the most current algorithm to tackle the problem is only shown to work with 3 catalogs. In this thesis, we extend this algorithm to allow for matching across a higher number of catalogs. In addition, we introduce a new algorithm that is more efficient and scales much better with large number of catalogs.
Description
Keywords
Optimization, statistics, astronomy, machine learning, decision theory framework, integer programming, stochastic optimization