Neural networks for classification and familiarity discrimination, with radar and sonar applications

ARTMAP neural models are real-time networks that blend Adaptive Resonance Theory (ART), fuzzy logic, evidence accumulation, and training set statistics for supervised category learning, pattern recognition, and prediction. This dissertation describes novel extensions to the ARTMAP family for enhanced performance on a variety of radar and sonar applications, including familiarity discrimination, data fusion, and real-time object recognition in a natural environment.; The recognition process involves both identification and familiarity discrimination. A novel extension to ARTMAP is developed in this dissertation which performs familiarity discrimination while retaining basic ARTMAP properties of fast learning and real-time operation. Experiments using simulated radar range profile data demonstrate the model's effectiveness, and techniques are developed which predict classification accuracy from training data.; A recent extension of ARTMAP employs an improved search algorithm and training set statistics to yield enhanced classification rates and reduced memory requirements. This dissertation describes the application of this extension (ARTMAP-IC) to the problem of simulated radar range profile classification. Performance is compared to that of fuzzy ARTMAP and k nearest neighbor systems.; A novel data fusion architecture based upon ARTMAP is also developed in this dissertation for the classification of sonar data of underwater objects. The system consists of a hierarchical arrangement of ARTMAP neural networks that use distributed category coding. System performance on a series of benchmark classification tasks is compared to other classification methodologies that use multi-layer perceptron networks. Optimal parameters are identified via a validation procedure based on training set data.; Finally, this dissertation describes a robotics application that employs ARTMAP to recognize objects in real time using the Polaroid sonar sensors found on many mobile robots. Ultrasonic echoes are collected from a variety of objects in a typical indoor environment, sampled, and transformed into the frequency domain. The power spectrum of the echoes is used as input to an ARTMAP network, which is trained to recognize the objects independently of their distance or orientation. Results demonstrate that this inexpensive hardware/software configuration can be the basis for a successful real-time object recognition system.