PREDICTING FUTURE DESTINATIONS OF TACTICAL UNITS

Abstract

The purpose of this thesis is to apply machine learning techniques towards predicting the future destinations of tactical units that move in a known road network. These units are modeled after standard field artillery batteries. Each battery is made up of eleven vehicles: four launcher vehicles, four reloading vehicles, two support vehicles, and one command control vehicle. Data was generated by the Modeling Virtual Environments and Simulation (MOVES) institute at NPS. There are two study questions: Can machine learning models accurately predict the future destinations of tactical vehicles? What is an adequate level of prediction accuracy for use in tactical applications?Of the current machine learning techniques, we use random forests and neural networks for destination prediction. Overall, our random forest achieves 38.9 percent prediction accuracy while our neural network achieves 43.2 percent prediction accuracy. There are four immediate directions for future research following this thesis. They are further investigation of prediction modeling, using data with measurement error collected on irregular time intervals, modeling with real world data, and multi-domain modeling.