Transitioning from Human to Agent-Based Role-Players for Simulation-Based Training

摘要

In the context of military training simulation, "semi-automated forces" are software agents that serve as role players. The term implies a degree of shared control - increased automation allows one operator to control a larger number of agents, but too much automation removes control from the instructor. The desired amount of control depends on the situation, so there is no single "best" level of automation. This paper describes the rationale and design for Trainable Automated Forces (TAF), which is based on training by example in order to reduce the development time for automated agents. A central issue is how TAF interprets demonstrated behaviors either as an example to follow specifically, or as contingencies to be executed as the situation permits. We describe the behavior recognizers that allow TAF to produce a high-level model of behaviors. We assess the accuracy of a recognizer for a simple airplane maneuver, showing that it can accurately recognize the maneuver from just a few examples.