Agent Smith: Towards an evolutionary rule-based agent for interactive dynamic games

Abstract

The goal of this project is to develop an agent to play the first-person shooter game Unreal Tournament 2004 , a fast-paced and dynamic environment that demands that the agent must be capable of making decisions quickly. An additional goal of this project is to explore evolutionary computation as a means for learning the rule sets used to control the game-playing agent. The agent's behavior is controlled by a rule-based system, which looks at multiple high-level conditions, such as whether the agent is weak, and determines a single high-level action, such as to head for the nearest known healing source. Using an evolutionary computation approach, in which the behavior is evolved over a number of generations, the agent learns increasingly better strategies for its environment. Through the work in this project, we are exploring several research questions, including the development of successful vocabulary of high-level conditions and actions for the rule set, the challenges of using the evolutionary process to hone a rule set, and the effects of using some expert knowledge in combination with the evolutionary process.