AI for automated combatants in a training application

We are trying to emulate the most realistic human-combat behaviors possible in a virtual environment. Our requirement for realism stems from our application: a combat training environment where the virtual arena needs to engage students in experiences that can carryover to "real life." For us, "realism" is focused on the behaviors of enemy combatants in the virtual world, and the definition of that reality needs to address questions like: Do the enemies hide from you? Where? How well? What weapons do they choose? How accurately do they aim? How fast can they shoot? Are they familiar with the tactics you and your team members use most often? Are they acting rationally given current conditions in terms of visibility, noise, and obstacles? In this paper we describe the system we are building: design choices, authoring concepts, and tools.