cjdarken At nps Dot edu

ABSTRACT: The information provided to a software agent fundamentally affects its behavior. It is a trivial observation that an agent cannot respond to an environmental stimulus of which it is not aware. In a similar vein, providing an agent with information that a human participant in the simulation would not be aware of in the same circumstances may result in inaccurate agent behavior. In 3D virtual simulations, the most basic information provided to an agent concerns what battlefield entities that they can see. The standard approach used in 3D simulations with high visual fidelity, such as video games, is to use a line-of-sight (LOS) trace between entities to determine if they can see each other. LOS is arguably a very poor model of target detection, particularly in its failure to take camouflage, smoke, and darkness into account. In this paper, we describe a new approach that is an adaptation of a standard target acquisition model, to the domain of high visual fidelity simulations rendered on conventional graphics hardware. The new approach avoids the problems of LOS. We describe two variants of the approach, compare their predictions to human performance, and characterize their remaining deficiencies. 1