The small unit tactical trainer (SUTT) is under development for the Marine Corps by the Naval Air Warfare Center, Training Systems Division (NAWCTSD). It is a virtual world simulator for training Marine rifle squads in military operations in urban terrain (MOUT) clearing (Reece, 1996). The virtual world in the trainer is populated by computer-controlled hostiles (CCHs) developed by the Institute for Simulation and Training at the University of Central Florida. The CCHs behave as smart adversaries and evade and counterattack friendly forces. The key state variables include low-level facts (e.g., soldier position, heading, posture), situation knowledge (e.g., threats, current target), and the current task. Perception is based on simple visual and aural detection and identification of enemy forces. Situation awareness involves both visible and remembered threats.
Most of the activity is reflexive. The model has an action selection component based on hierarchical task decomposition, starting with top-level-tasks, e.g., engage enemy, seek cover, watch threat, look around, and run away. Situation dependent rules propose tasks to perform. Rule priorities allow specification of critical, normal, and default behavior. Tasks may be deliberately proposed with the same priority level, allowing random selection to choose a task according to predefined weights. This mechanism supports variation in behavior that obeys a prior probability distribution (e.g., probability of fight or flight reaction can be set for a CCH). A new task may be selected each time frame as the situation changes. Proposed, but not implemented, architecture extensions would allow the selection of tasks that would serve more than one top-level goal or task simultaneously. Other models under development as part of this program include one for visual detection and another for hearing (Reece and Kelly, 1996).
IFOR models have been created using the Soar architecture to model the combat behavior of fixed- and rotary-wing pilots in combat and reconnaissance missions. (Details on the Soar architecture are presented in Chapter 3.) IFORs are very large expert systems—they use encoded knowledge as a basis for action and problem solving. The Soar architecture was originally devised to support the study of human problem solving behavior. It incorporates many of the concepts of artificial intelligence, and its main feature is the use of production rules as the means to link an initial condition (a stimulus) to a particular response. Although Soar is capable of learning, this function has not been exercised.
To meet the objectives of simulating intelligent forces, specific contexts were needed. In the IFOR framework, adaptations for both fixed-wing attack (FWA)-Soar and rotary-wing attack (RWA)-Soar air operations were developed.