NEURAL NETWORK TECHNIQUES FOR PRACTICAL APPLICATIONS TO TIME SERIES PREDICTION

ALTHOUGH IT IS WIDELY KNOWN THAT NEURAL NETWORKS HAVE, IN SPECIFIC CASES, DEMONSTRATED PERFORMANCE SUPERIOR TO TRADITIONAL METHODS OF TIME SERIES PREDICTION, THERE IS STILL MUCH TO BE LEARNED WITH RESPECT TO THEIR APPLICATION TO PROBLEMS WHERE THE DYNAMICS OF THE UNDERLYING SYSTEM ARE NON-STATIONARY. THIS PROPOSAL ADDRESSES A SPECIFIC APPLICATION OF NEURAL NETWORK TECHNOLOGY TO A FORECASTING PROBLEM WITH BROAD MILITARY IMPLICATIONS. THE PROPOSED PROBLEM, THAT OF FORECASTING THE MOTION OF A SHIP DUE TO OCEAN-WAVE ACTIVITY, HAS A NUMBER OF FEATURES THAT PLACE A HIGH VALUE ON ITS SOLUTION AT THIS TIME. SINCE THE PROBLEM IS PRESENTLY UNDER STUDY USING ANOTHER PREDICTION METHODOLOGY, IT IS PROPOSED TO: (1) DERIVE A NEURAL NETWORK THAT IS APPROPRIATE TO THE SHIP MOTION FORECASTING PROBLEM, (2) COMPARE AND CONTRAST THE PERFORMANCE CHARACTERISTICS OF THE ALTERNATIVE FORECASTING TECHNIQUE WITH THOSE OF THE NEURAL NETWORK APPROACH UTILIZING THE SAME SET OF AIRCRAFT CARRIER MOTION DATA IN A UNIFIED EXPERIMENTAL DESIGN, (3) DEVELOP PERFORMANCE METRICS AND VALIDATION TECHNIQUES FOR OTHER NEURAL NETWORK APPROACHES TO TIME SERIES FORECASTING, AND (4) LAY THE GROUNDWORK FOR A PHASE II EFFORT COMPRISING ADVANCED TECHNIQUES FOR FORECASTING FLUIDIC WAVE MOTION USING THE BEST PROPERTIES OF EITHER TECHNIQUE OR, IF APPROPRIATE, A COMBINED APPROACH. SOME PRELIMINARY DATA ARE PRESENTED, SUGGESTING THAT THE NEURAL NETWORK APPROACH EXHIBITS SUPERIOR FORECASTING CAPABILITIES OVER OTHER TECHNIQUES. ANTICIPATED BENEFITS/POTENTIAL APPLICATIONS - THE ANTICIPATED BENEFITS OF THE PROPOSED FORECASTING SCHEME WILL BE IMPROVED SAFETY AND PERFORMANCE OF A VARIETY OF SHIP-BOARD AIRCRAFT OPERATIONS BY THE U.S. COAST GUARD, MARINES, AND NAVY. POTENTIAL COMMERCIAL APPLICATIONS INCLUDE: ROTARY WING OPERATIONS, ECONOMIC AND LOGISTICAL MODELS, CLIMATE AND WEATHER PREDICTIONS, AND INVENTORY CONTROL FOR LARGE COMPANIES.