Thesis (Ph.D)--Virginia Commonwealth University, 2009. / Prepared for: Dept. of Human Genetics. Title from title-page of electronic thesis. Bibliography: leaves 117-151.

Thesis--PlanB (M.S.)--University of Wisconsin--Stout, 2003. / Includes bibliographical references.

Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 2009. / Source: Dissertation Abstracts International, Volume: 70-06, Section: B, page: 3764. Adviser: Nikolaos Sahinidis. Includes bibliographical references (leaves 109-119) Available on microfilm from Pro Quest Information and Learning.

Not available / text

No description available.

No description available.

No description available.

Data mining is a semi-automated technique to discover patterns and trends in large amounts of data and can be used to build statistical models to predict those patterns and trends. One type of prediction model is a classifier, which attempts to predict to which group a particular item belongs. An important binary classifier, the Support Vector Machine classifier, uses non-linear optimization to find a hyperplane separating the two classes of data. This classifier has been reformulated as a linear program and as a pure quadratic program. We propose two modeling extensions to the Support Vector Machine classifier. The first, the Linearized Proximal Support Vector Machine classifier, linearizes the objective function of the pure quadratic version. This reduces the importance the classifier places on outlying data points. The second extension improves the conceptual accuracy of the model. The Integer Support Vector Machine classifier uses binary indicator variables to indicate potential misclassification errors and minimizes these errors directly. Performance of both these new classifiers was evaluated on a simple two dimensional data set as well as on several data sets commonly used in the literature and was compared to the original classifiers. These classifiers were then used to build a model to detect collusion in online gambling. Collusion occurs when two or more players play differently against each other than against the rest of the players. Since their communication cannot be intercepted, collusion is easier for online gamblers. However, collusion can still be identified by examining the playing style of the colluding players. By analyzing the record of play from online poker, a model to predict whether a hand contains colluding players or not can be built. We found that these new classifiers performed about as well as previous classifiers and sometimes worse and sometimes better. We also found that one form of online collusion could be detected, but not perfectly.

In this dissertation we address large, multiple-mode, resource-constrained project scheduling problems with the objective of minimizing makespan. After noting that projects often fail and new research is needed, we provide the formal definition of the resource-constrained project scheduling problem and review the existing literature. We then introduce a new model based on digital electronics. We conceptualize our model using agent technology and discuss it as extension of existing models with more representational power. We also describe how our model supports distributed planning. After implementing our model, we conduct two computational studies. In the first, we develop two agent types: basic and enhanced where the enhanced agent is more sophisticated in selecting an activity execution mode. We apply these agents to the scheduling of 500 instances of a small project originally published by Maroto and Tormos (1994). We evaluate the performance of the agents in conjunction with their use of eight heuristic prioritization rules: shortest and longest processing time, fewest and most immediate successors, smallest and greatest resource demand, earliest start time, and earliest due date. Our results show that enhanced agents consistently outperform basic agents while the results regarding priority rules were mixed. In the second computational study, we further develop our enhanced agents by providing still more sophisticated mode selection. We also evaluate static versus dynamic prioritization and two more priority rules: shortest and longest duration critical path. For this study we generated 2500, 5000, 7500, and 10000 activity projects. For each of these, we generated networks with complexities of 1.5, 1.8, and 2.1. For these twelve networks, we generated 20 problem instances for every possible combination of resource factor = 0.25, 0.50, 0.75, 1.0 and resource strength = 0.2, 0.5, 0.8. We graphically evaluated scheduling performance, computation times, and failure rates and conducted an extensive statistical analysis. We found that enhanced agents using shortest processing time priority consistently produced the shortest schedules. However, these agents fail more often than basic agents. We found that dynamic prioritization requires more computation time, but provides little improvement in scheduling performance. We conclude this work with suggestions for future research.

No description available.