Active Behavior Recognition in Beyond Visual Range Air Combat.

摘要

Accurately modeling uncontrolled agents (or recognizing their behavior and intentions) is critical to planning and acting in a multi-agent environment. However, behavior recognition systems are only as good as their observations. Here we argue that acting, even acting at random, can be a critical part of gathering those observations. Furthermore, we claim that acting intelligently via automated planning can significantly reduce the time it takes to confidently classify agent behaviors. We present a formalism and algorithm for integrated planning and recognition, as well as its implementation in a beyond visual range air combat simulator. We found that it yields better behavior recognition than non-integrated approaches. This provides evidence that behavior recognition is not just a necessary component of an intelligent agent, but that good behavior recognition requires intelligent acting.