USE OF NEURAL NETWORK TECHNOLOGY IN PREDICTING PERSONNEL ATTRITION

MILITARY LOSS FORECASTING IS AN IMPORTANT PROBLEM FOR SEVERAL REASONS. PERSONNEL LOSSES ARE COSTLY BECAUSE THEY INCREASE THE NUMBER OF PEOPLE WHO MUST BE RECRUITED AND TRAINED. MORE SERIOUS IS THE PROSPECT OF MILITARY MISSIONS DEGRADED BY MANPOWER SHORTAGES. THERE ARE SEVERAL DIFFERENT TYPES OF FORECASTING METHODOLOGIES IN USE AT THE PRESENT TIME. FOR EXAMPLE ORDINARY LEAST SQURES STEPWISE REGRESSION BASED ON INDIVIDUAL CHARACTERISTICS AND A VERSION OF ODDS-FOR-EFFECTIVENESS TABLES TO ESTIMATE ATTRITION PROBABILITY ARE USED BY AIR FORCE AND NAVY TO PREDICT LOSSES ON AN INDIVIDUAL BASIS. IN ALL THESE CASES THERE ARE SOME PREJUDICES OF THE RESEARCHERS BUILT INTO THE MODEL AND IT IS TIME TO STUDY MODELS WHICH LOOK ONLY AT THE DATA AND NOTHING ELSE. NEURAL NETWORK TECHNOLOGY HAS GIVEN US THIS PRIVILEGE OF CREATING MODELS WHICH TRAIN THEMSELVES ON THE DATA AND THERE BY REMOVING ALL THE ASSUMPTIONS BUILT INTO CURRENT FORECASTING TECHNIQUES. DARPA NEURAL NETWORK STUDY RECOMMENDS GMDH AS ONE OF THE MOST USEFUL TECHNIQUES IN CLASSIFICATION AND MODELING. WE HAVE DEVELOPED A TECHNIQUE FOR FURTHER REFINEMENT ON THE GMDH TECHNIQUE WHICH INCORPORATES DIFFERENT TYPES OF LEARNING TECHNIQUES, ENABLING US TO CREATE A NEURAL NETWORK ARCHITECTURE WITH OPTIMAL COMPLEXITY. WE ALSO PROPOSE AN ALTERNATE METHOD TO UTILIZE THE NEURAL NETWORK TECHNOLOGY DEVELOPED USING LARS AND INTELEHEDRON DEVELOPED BY OUR COLLABORATORS FROM GTE, IN THIS STUDY.