TARGET RECOGNITION TRACKING AND RESPONSE WITH NEURAL NETWORK KALMAN-BUCY FILTERS

THE PROPOSED RESEARCH WILL INCORPORATE THE PARAMETRIC AVALANCHE NEURAL NETWORK ARCHITECTURE INTO A MAXIMUM LIKELIHOOD DETECTION AND TRACKING ALGORITHM FOR USE IN AUTONOMOUS WEAPON SYSTEMS. THE PARAMETRIC AVALANCHE WILL PERFORM O(1) SEARCH IN ITS STORED PATTERN FEATURES TO IDENTIFY THE TARGET AND ITS MOTION EQUATIONS, AFTER WHICH IT WILL TRACK THE TARGET. A SECOND PARAMETRIC AVALANCH WILL USE THE STATE ESTIMATES FROM THE FIRST, TOGETHER WITH A REFERENCE TRAJECTORY SUPPLIED PERHAPS BY AN EXPERT SYSTEM, TO GENERATE CONTROL SIGNALS FOR GUIDANCE OF THE AUTONOMOUS WEAPON SYSTEM. THE PARAMETRIC AVALANCHE IS A RECURRENT TWO-LAYER NEURAL NETWORK WHICH EMPLOYS A PROPRIETARY TECHNIQUE FOR UNSUPERVISED LEARNING OF AN INTERNAL MODEL OF OBSERVED DYNAMICAL SYSTEMS AND SUBSEQUENTLY EMPLOYING THAT MODEL IN A CONTINUOUS BAYESIAN ESTIMATOR (GENERALIZING THE KALMAN-BUCY FILTER). THE ARCHITECTURE EMPLOYS A NOVELTY FILTER TO WHITEN THE PREDICTABLES IN THE SCENE, AND IT NEEDS NO A-PRIOR KNOWLEDGE OF STRUCTURE IN THE SCENE, SO THAT IT CAN IN PRINCIPLE INTEGRATE SENSOR INPUTS FROM VARIOUS SOURCES. TRACKING TAKES PLACE SUBLIMINALLY UNTIL CONTINUOUS LIKELIHOOD REINFORCEMENT TRIGGERS DETECTION.