| The Rhagoletis pomonella species complex in the United States consists of sibling species and host races including four species: R. pomonella (Walsh), R. mendax Curran, R. zephyria Snow, and R. cornivora Bush, at least two un-described species: sparkleberry fly and flowering dogwood fly; and two indeterminate populations: plum fly and mayhaw fly; and at least two host races, R. pomonellaHR/HAW and R. pomonella HP/APPLE.; I hypothesize that there is hidden biological information in the wing vein structure and banding patterns in the pomonella species group that can be used to distinguish them. I also hypothesize that allozymes may be useful to the classification of this species group. To test the hypotheses I used engineering methods, including pattern recognition, fuzzy logic, neural networks and data mining techniques, to explore the experimental datasets, and build models and discovery knowledge from them and thus help classification of the pomonella species group.; The wing dataset was divided into male and female subsets, and from each models were built. The classifying models consist of Fisher linear discriminant functions (FLDF), probability neural networks (PNN), membership functions built from Genetic Fuzzy Clustering (GFC) and fuzzy rule bases extracted from Neuro-Fuzzy computing model. The genetic algorithm for feature selection (GAFS) was used to optimize all the models. Among them the optimized FLDF models showed the highest performance (85% success rate). PNN models have the capability of memorizing the prototypic/reference patterns.; For the allozyme dataset a small set of database management functions were designed and implemented to manipulate the data. The Bayes decision theory was used to build the allozyme classifier. The probabilities of allozyme phenotypes on each locus of different taxon were estimated by data mining the dataset. The unique phenotypes and association rules for each taxon were also mined and combined with the Bayesian classifier. The test performance reaches 79% on average. Among the nine taxa R. zephyria, and R. cornivora can be identified without error due to their unique phenotypes.; Finally an intelligent hybrid classification system was built to integrate each of these classification tools. A small set of complete new wing morphometric data was used to test the predictions and the results were satisfactory. The hypotheses were supported. The results have implications for agricultural production and quarantine issues and could be helpful in devising a classification system for rapid identification of invasive species at ports of entry. (Abstract shortened by UMI.)... |
