This research aims at learning the behavior of operators or role players in military simulators based on observed behavior. The paper describes the principle of so-called imitation learning that we used to learn behavior based on observations of players. We present an overview of the existing examples of the current application of data-driven behavior modeling in the military domain. We apply imitation learning to the army ground based air defense system (AGBADS) operator behavior, for which real-life data (from training events) was available. The data that is used in the learning process originates from the 2021 air-defense exercise Joint Project Optic Windmill (JPOW). After extensive pre-processing, multiple behavior models were constructed that can be used to provide fire control solutions (when to fire, the salvo size and which launcher to use) for a ground based air defense system given a tactical situation.
We create explainable models using decision trees, which can help during the after action review process to give trainees insight in their actions when training as an AGBADS operator. Using less explainable but more precise models, we can create replicas of behavior of an AGBADS, which can be used to simulate the AGBADS in training. Furthermore, the models could be used to aid decision support for AGBADS operators. We achieved quite reasonable performances with regards to the models, especially considering that we had to work with a discrepancy between logged data and the perceived situation as experienced by the operator, making it hard to create an accurate feature representation of the situation.
Keywords
AIR AND MISSILE DEFENSE,BEHAVIOR MODELING,DATA,MACHINE LEARNING,RAPID MODELING
Additional Keywords