Global Optimization of Finite Mixture Models
| dc.contributor.advisor | Fu, Michael | en_US |
| dc.contributor.advisor | Jank, Wolfgang | en_US |
| dc.contributor.author | Heath, Jeffrey Wells | en_US |
| dc.contributor.department | Applied Mathematics and Scientific Computation | en_US |
| dc.contributor.publisher | Digital Repository at the University of Maryland | en_US |
| dc.contributor.publisher | University of Maryland (College Park, Md.) | en_US |
| dc.date.accessioned | 2007-09-28T14:56:59Z | |
| dc.date.available | 2007-09-28T14:56:59Z | |
| dc.date.issued | 2007-05-31 | en_US |
| dc.description.abstract | The Expectation-Maximization (EM) algorithm is a popular and convenient tool for the estimation of Gaussian mixture models and its natural extension, model-based clustering. However, while the algorithm is convenient to implement and numerically very stable, it only produces solutions that are locally optimal. Thus, EM may not achieve the globally optimal solution in Gaussian mixture analysis problems, which can have a large number of local optima. This dissertation introduces several new algorithms designed to produce globally optimal solutions for Gaussian mixture models. The building blocks for these algorithms are methods from the operations research literature, namely the Cross-Entropy (CE) method and Model Reference Adaptive Search (MRAS). The new algorithms we propose must efficiently simulate positive definite covariance matrices of the Gaussian mixture components. We propose several new solutions to this problem. One solution is to blend the updating procedure of CE and MRAS with the principles of Expectation-Maximization updating for the covariance matrices, leading to two new algorithms, CE-EM and MRAS-EM. We also propose two additional algorithms, CE-CD and MRAS-CD, which rely on the Cholesky decomposition to construct the random covariance matrices. Numerical experiments illustrate the effectiveness of the proposed algorithms in finding global optima where the classical EM fails to do so. We find that although a single run of the new algorithms may be slower than EM, they have the potential of producing significantly better global solutions to the model-based clustering problem. We also show that the global optimum matters in the sense that it significantly improves the clustering task. Furthermore, we provide a a theoretical proof of global convergence to the optimal solution of the likelihood function of Gaussian mixtures for one of the algorithms, namely MRAS-CD. This offers support that the algorithm is not merely an ad-hoc heuristic, but is systematically designed to produce global solutions to Gaussian mixture models. Finally, we investigate the fitness landscape of Gaussian mixture models and give evidence for why this is a difficult global optimization problem. We discuss different metrics that can be used to evaluate the difficulty of global optimization problems, and then apply them to the context of Gaussian mixture models. | en_US |
| dc.format.extent | 2018506 bytes | |
| dc.format.mimetype | application/pdf | |
| dc.identifier.uri | http://hdl.handle.net/1903/7179 | |
| dc.language.iso | en_US | |
| dc.subject.pqcontrolled | Applied Mechanics | en_US |
| dc.subject.pqcontrolled | Operations Research | en_US |
| dc.subject.pqcontrolled | Statistics | en_US |
| dc.title | Global Optimization of Finite Mixture Models | en_US |
| dc.type | Dissertation | en_US |
Files
Original bundle
1 - 1 of 1