| This thesis is concerned with processing data from multiple sensors of an electronic nose, with the goal of determining which sensors are the most useful in discriminating one odor from another. The most significant obstacle to selecting the best sensors is the extremely small number of available examples, together with the very large number of features. To ensure that the results are as accurate as possible, several aspects of the statistical pattern recognition process were investigated first, in order to find out which methods work best under these conditions. Once the best low-level procedures were identified, the sensor selection process was analyzed.; The results indicate that a combination of the Resubstitution and Leave-One-Out estimators provides the most accurate non-stochastic estimation of classifier success when there are fewer examples than features. Those feature selection algorithms which most often provided the best features subsets are also identified. In addition, many different tests show that to have any confidence in the results, there must be at least twice as many examples as features, and five or ten times would be better. However, some small-sample effects do not completely disappear until there are more than forty or fifty examples per feature. |
