GRUE : an architecture for agents in games and other real-time environments
This thesis presents an architecture, which we call GRUE, for intelligent agents in real-time dynamic worlds. Such environments require agents to be able to flexibly adjust their behaviour to take into account changes in the environment or other agents’ actions. Our architecture is based on work done in robotics (Nilsson, 1994; Benson and Nilsson, 1995; Benson, 1996), which also deals with complex, dynamic environments. Our work focuses on goal arbitration, the method used by the agent to choose an appropriate goal for the current situation, and to re-evaluate when the situation changes. In the process, we have also developed a method for representing items in the environment, which we call resources, in terms of their properties. This allows the agent to specify a needed object in terms of required properties and use available objects with appropriate properties interchangeably. We show that the GRUE architecture can be used successfully in both a typical AI test bed and a commercial game environment. In addition, we have undertaken to experimentally test the effects of the features included in our architecture by comparing agents using the standard GRUE architecture to agents with one or more features removed and find that these features do improve the performance of the agent where expected.