Properties of learning of the Fuzzy ART neural network and improvements of the generalization performance of the Fuzzy ARTMAP neural network.

摘要

Pattern classification is a key element to many engineering solutions. Sonar, radar, seismic, and diagnostic applications all require the ability to accurately classify data. Control, tracking and prediction systems will often use classifiers to determine input-output relationships. Because of this wide range of applicability, pattern classification has been studied a great deal. A number of desirable properties that a pattern classifier should possess are listed below. Property 1: On-Line Adaptation, Property 2: Non-Linear Separability, Property 3: Short Training Time, Property 4: Soft and Hard Decisions, Property 5: Verification and Validation, Property 6: Independence from Tuning Parameters, Property 7: Nonparametric Classification, and Property 8: Overlapping Classes. A neural network classifier that satisfies most of these properties is Fuzzy ARTMAP (e.g., properties 1, 2, 3, 4, 5, 7). In the first part of this dissertation we analytically prove the short training time property of a Fuzzy ART Variant. Fuzzy ART is an important component of the Fuzzy ARTMAP neural network, and proving Fuzzy ART's short training time capability is the first step in proving Fuzzy ARTMAP's short training time capability (Property 3). In the second part of this dissertation we introduce a variation of the Fuzzy ARTMAP network, called Fuzzy ARTVar, whose performance is better than the Fuzzy ARTMAP's performance. Fuzzy ARTVar achieves better performance than Fuzzy ARTMAP by identifying network weights that represent the data more accurately. Furthermore, Fuzzy ARTVar's performance is independent of the tuning of parameters (Property 6), in contrast to other variations of Fuzzy ARTMAP that have appeared in the literature and improved its performance but depend on certain network parameters. Finally, the third part of the dissertation tackles an old nemesis of Fuzzy ARTMAP, which is its performance dependence on the order of the training pattern presentation (violation of Property 6). To remedy this problem, we introduce a systematic procedure that identifies the order according to which training patterns are to be presented in Fuzzy ARTMAP. The resulting algorithm, designated as Ordered Fuzzy ARTMAP, exhibits a performance that is better than the average Fuzzy ARTMAP's performance (average of a fixed number of Fuzzy ARTMAP's performances corresponding to random orders of training pattern presentations) and occasionally as good as, or better than the maximum Fuzzy ARTMAP's performance (maximum of a fixed number of Fuzzy ARTMAP's performances corresponding to random orders of training pattern presentations).