For computer generated forces to be useful in training environments, they must exhibit a broad range of skill levels, competency at their assigned missions, and comply with current doctrine. Because of the rapid rate of change in Distributed Interactive Simulation and the expanding set of performance objectives for any computer generated force, the system must also be modifiable at reasonable cost and incorporate mechanisms for learning. The requirements pose an intricate set of challenges because the system must satisfy reasoning and fidelity requirements as well as performance requirements. To address these circumstances, we developed a set of general requirements for aircraft computer generated forces and used them to guide our specification of a generalized architecture for aircraft computer generated forces. In this paper, we present a component-wise decomposition of the system and describe the structure of the major components of the computer generated force decision mechanism. We illustrate the application of this architecture by presenting its application to the design of an aircraft computer generated force, the Automated Wingman.