This paper reports a means for transferring behavioral observations captured in operational settings to training simulation. The goal is to enable trainees in a simulation environment to experience training scenarios that mimic observed operational conditions, including specific tactics, techniques and procedures as practiced. The emerging capability generates simulation-training scenarios that replicate events observed in the original setting while also supporting trainee interaction (i.e., not just a scripted replay). The approach is designed to be both simulation agnostic and computationally tractable. Thus, it should be applicable to many simulations and user communities. The paper reports feasibility as implemented within a widely-used DoD distributed simulation environment.
To achieve scenario-generation capability, artificial intelligence and machine learning are used to develop models of the behavioral capabilities of a target simulation and how they can be instantiated and sequenced to produce behaviors, events, and outcomes. Using the inherent capabilities of a simulation itself offers two notable advantages. First, it significantly reduces computational requirements, transforming behavior generation challenges into far simpler recognition problems. Second, it enables interactivity and variation during training by drawing on the underlying simulation’s behavioral capabilities. These simplifications make realistic-scale scenario generation tractable.
We describe scenario-generation goals, multiple approaches that we and others have investigated, and results of empirical verification demonstrating a recent approach is functionally sufficient and computationally tractable. In this latest version, captured data is segmented into entity-centric events (e.g., an entity turning “hot” onto an adversary). These event sequences are compared to “behavior signatures” generalized from systematic generation of simulation data. Using algorithms adapted from genetic sequencing, an “event matching” algorithm identifies behavioral signatures that closely align with sequences observed in the original exercise. The resulting simulation scenarios closely reproduce the sequence of events in the original data, but also allow interaction and deviation from the original sequence.