Fuzzy ARTMAP has to date been applied to a variety of automatic target recognition tasks, including radar range profile classification. In simulations of this task, it has demonstrated significant compression compared to k-nearest-neighbor classifiers. During supervised learning, match tracking search allocates memory based on the degree of similarity between newly encountered and previously encountered inputs, regardless of their prior predictive success. Here we investigate techniques that buffer reset based on a category's previous predictive success and thereby substantially improve the compression achieved with minimal loss of accuracy.
展开▼
