In a real world pattern recognition application a user cannot assess the performance of a classifier on an unlabeled data set. Classifiers cannot give their best performance because they require user-controlled parameters. As a Solution, a Sequential Function Approximation (SFA) method has been' developed for classification that determines the values of the control parameters during learning. In this dissertation, experiments were carried out on real world data sets where SFA, using only the training subset, had comparable performance to a number of other popular classification schemes whose user-defined parameters were optimized utilizing the entire data set. By the statistical significance of the results it was concluded at 95% confidence that the performance of SFA will be equivalent or significantly better than those of the other popular classification tools. After establishing SFA as a proper classification tool in this dissertation, it is applied to a US Navy flight test problem. The current problem at hand is to predict pilot ratings from HH-60H Sea-Hawk helicopters based on 369 at sea take-off and landing DI tests. Least significant inputs with respect to classification were pointed out with the potential of accelerating through the DI test matrix. And finally an effort was made to give the DI test pilots an estimate of how many tests were necessary to be conducted before generating enough data for the SFA classification tool to satisfactorily learn.
| Identifer | oai:union.ndltd.org:RICE/oai:scholarship.rice.edu:1911/20540 |
| Date | January 2007 |
| Creators | Srivastava, Ankur |
| Contributors | Meade, Andrew J., Jr. |
| Source Sets | Rice University |
| Language | English |
| Detected Language | English |
| Type | Thesis, Text |
| Format | 78 p., application/pdf |
Page generated in 0.0021 seconds