APPLICATION OF ADAPTIVE AND MODEL-BASED DIAGNOSIS TECHNOLOGY TO SATELLITE SYSTEMS

MUCH RECENT WORK HAS BEEN FOCUSED ON ELUCIDATING THE POTENTIAL BENEFITS OF USING FUZZY LOGIC AND ADAPTIVE NETWORK DECISION MAKING METHODS. ADAPTIVE METHODS, LIKE BACK-PROPAGATION TAUGHT NEURAL NETWORKS, CAN MAKE IT POSSIBLE TO BUILD COMPLEX DECISION MAKING SYSTEM WHICH AVOID THE "EXPERT SYSTEMS" PROBLEM OF MANUALLY ACQUIRED KNOWLEDGE "CAPTURE." INSTEAD, THESE ADAPTIVE METHODS CAPTURE KNOWLEDGE BY "LEARNING" FROM EXAMPLE DATASETS WHICH RELATE INPUTS TO DESIRED OUTPUTS (SUCH AS THE DECISION OF A SEISMIC EXPERT). UNFORTUNATELY, VERY LITTLE IS KNOWN ABOUT HOW TO STRUCTURE NEURAL NETWORKS OPTIMALLY FOR PARTICULAR DECISION MAKING TASKS, HOW MUCH TRAINING IS NECESSARY FOR ADEQUATE NETWORK PERFORMANCE, AND HOW TO VALIDATE TRAINED NETWORK PERFORMANCE RELATIVE TO SPECIFICATIONS. WE PROPOSE IMPLEMENTING A SUPERCOMPUTER NETWORK TRAINING, DECISION MAKING, AND VALIDATION TESTBED WHICH WILL INCORPORATE A VARIETY OF POPULAR ADAPTIVE NETWORK MODELS, AND WILL PROVIDE A MEANS FOR TESTING SOME PROPOSED CONCEPTS FOR NETWORK SYNTHESIS WHICH MAY MAKE DESIGNING ADAPTIVE METWORKS TO A SPECIFICATION MORE EASILY DONE FOR IMAGE DATA INTERPRETATION.