Mike van Lent and John Laird University of Michigan AI Lab 1101 Beal Av. Ann Arbor, MI 48109-2110 vanlent@umich.edu laird@umich.edu As computer games become more complex and consumers demand more sophisticated computer controlled agents, developers are required to place a greater emphasis on the artificial intelligence aspects of their games. One source of sophisticated AI techniques is the academic artificial intelligence community. We would like to propose that a central topic for discussion at the 1999 AAAI Spring Symposium on Artificial Intelligence and Computer Games should be how techniques from the academic community can be best utilized by the commercial game community. Recent work by our group at the University of Michigan Artificial Intelligence Lab has addressed this topic by creating an interface between the Soar artificial intelligence architecture and the commercial computer games Quake 2 and Descent III. As described in John Laird's invited talk at the 1998 Computer Game Developer's Conference our experience developing intelligence air combat agents for DARPA has led to advanced AI techniques and systems which are very applicable to computer games. Our approach, based on the system created for the DARPA project, has been to build a two part "behavior engine" that is used to make the computer controlled agents behave intelligently. The first part of the engine is a large knowledge base of game independent behaviors programmed as production rules (5200 rules for the DARPA project). This behavior knowledge base doesn't include game specific information but instead focuses on goals, strategies and tactics which apply to any game within a genre. For example, the circle strafing tactic would be a major component of the behavior knowledge base for games such as Quake, Quake 2, Doom, Sin, Unreal and many more. To apply the behavior engine to a specific game a small amount of game dependent information is added which would allow the circle strafing tactic to be applied differently according to the game dynamics. Thus the same circle strafing behavior could be used in Descent III with minor modifications to allow circling in two dimensions. The second part of the behavior engine is the Soar artificial intelligence architecture which serves as the inference engine to generate actions based on the behavior knowledge base, the agent's goals and internal memory and the current situation. The Soar architecture is the result of 15 years of research in the fields of artificial intelligence and cognitive psychology at various research universities. Although Soar has filled a wide variety of roles in the current context it is best described as an artificial intelligence specific programming language which can efficiently using very large amounts of knowledge. Soar also supports a number of advanced AI features such as hierarchical operator decomposition, goal directed reasoning and simple forms of learning. Exploration of the commercial applications of the Soar architecture has only recently begun. One large potential application is the generation of behavior in computer games and simulations. One obvious advantage of the behavior engine approach is the reusability of the game independent behavior knowledge. Additionally the operator-based nature of the knowledge base, as required by Soar, is very modular which allows programmers to easily mix and match strategies, tactics and goals as appropriate for their game. Some advantages specific to the Soar architecture include reactivity, goal directed reasoning, hierarchical behavior decomposition and some simple forms of agent learning. On a reasonable machine Soar is capable of controlling 6 to 8 agents at 10 cycles per second which results in reaction times similar to a human player. Goal directed reasoning and hierarchical behavior decomposition have been found to be so useful that they have become built in aspects of the Soar architecture. Finally, the Soar architecture supports a simple form of learning to speed up behavior generation in previously encountered situations.