INTEGRATION OF KNOWLEDGE-BASED SYSTEMS AND NEURAL NETWORKS FOR INTELLIGENT SENSORIMOTOR CONTROL

THE OBJECTIVE OF THIS RESEARCH IS TO DEVELOP HIERARCHICAL STRUCTURES FOR THE INTELLIGENT CONTROL OF HIGHLY REDUNDANT, DEXTEROUS ROBOTIC MANIPULATORS. USING MODELS OF HUMAN MOTOR SKILL ACQUISITION TO GUIDE THE INTEGRATION OF KNOWLEDGE-BASED SYSTEMS AND ARTIFICIAL NEURAL NETWORKS, THE PROPOSED CONTROL SCHEME PARALLELS THE TRAINING OF AN ATHLETEBY A COACH, WHEREBY THE ROBOT LEARNS THROUGH EXPERIENCE HOW TO PERFECT TASKS INITIALLY SPECIFIED IN A HIGH-LEVEL TASK LANGUAGE. KNOWLEDGE-BASED SYSTEM COMPONENTS ENCODE NEURAL NETWORK LEARNING STRATEGIES, AND SKILL ACQUISITION IS ASSOCIATED WITH THE SHIFT FROM A PREDOMINATELY KNOWLEDGE-BASED REPRESENTATION OF CONTROL TO A PREDOMINATELYNETWORK-BASED FORM. THIS REPRESENTATIONAL SHIFT AFFECTS THEALLOCATION OF COMPUTATIONAL RESOURCES, THE FOCUSING OF ATTENTION, AND THE ABILITY TO ADAPT. PHASE I OF THE PROPOSAL WILL DETERMINE THE FEASIBILITY OF APPLYING THIS CONTROL TECHNIQUE TO THE DIFFICULT PROBLEM OF LEARNING TO POSITION THE LOADED ENDPOINT OF A REDUNDANT MANIPULATOR SUBJECT TO MULTIPLE GOALS AND CONSTRAINTS. UTILIZING APPLICABLE PLANNING ALGORITHMS, RULE-BASED CONTROL SYSTEM COMPONENTS WILL INITIALLY TRAIN, THEN SHIFT CONTROL TO, NEURAL NETWORK COMPONENTS. NETWORK PERFORMANCE WILL THEN BEREFINED THROUGH REINFORCEMENT LEARNING AND ON-LINE OPTIMIZATION. THE RESULTING CONTROL TECHNIQUE, COMBINING CONVENIENCE IN DESIGN WITH EFFICIENCY IN IMPLEMENTATION, WILL LAY THE FOUNDATION FOR THE DEVELOPMENT OF ROBOTIC SYSTEMS CAPABLE OF ACQUIRING THE SKILLS NECESSARY TO ACCOMPLISH COMPLEX TASKS IN UNSTRUCTURED ENVIRONMENTS.