Constructing Game Agents Through Simulated Evolution (2016)
Computer game worlds are often inhabited by numerous artificial agents, which may be helpful, neutral, or hostile toward the player or players. Common approaches for defining the behavior of such agents include rule-based scripts and finite state machines. However, agent behavior can also be generated automatically using evolutionary computation (EC). EC is a machine-learning technique that can be applied to sequential decision-making problems with large and partially observable state spaces, like video games. EC can create individual agents or teams, and these agents can be opponents or companions of human players. Agents can also be evolved to play games as a human would, in order to test the efficacy of EC techniques. EC can even create game artifacts besides agents, such as weapons. The reason EC is so flexible is that it requires little domain knowledge compared to traditional approaches. It is also capable of discovering surprising and effective behavior that a human expert would not think to program. If applied intelligently, this approach can even adapt to humans in a manner that keeps providing interesting and novel experiences for players. This article focuses mostly on discovering effective opponent behavior (since that is the focus of most research), although examples of other applications are also given when appropriate.
In Newton Lee, editors, Encyclopedia of Computer Graphics and Games, 1--10, 2016. Springer.

Risto Miikkulainen Faculty risto [at] cs utexas edu
Jacob Schrum Ph.D. Alumni schrum2 [at] southwestern edu