A new LMS-based algorithm for rapid adaptive classification in dynamic environments: theory and preliminary results

An algorithm has been developed for supervised adaptive classifications with rapid incremental learning characteristics in dynamic environments. The algorithm outperforms the Widrow-Hoff algorithm in applications when: (1) real-time response is required; (2) the classifier is subjected to repeated to train-then-apply cycles; and (3) the data point distributions can change dynamically. The algorithm is based on the concept of ´balancing´ actions during each training cycle. Overhead for the improved performance is not excessive, and other strategies are devised to reduce the overall computational complexity, so that the algorithm is simple to implement either in software or hardware. The algorithm is benchmarked against the Widrow-Hoff algorithm to work as a clutter rejection unit in a target tracking system. Performance is shown for three different sets of forward-looking infrared radar data. In all cases, the proposed algorithm outperformed Widrow-Hoff in terms of reduced misclassifications.<>