AI for Controlling Non-player Characters

The founding fathers of computer science and information theory (Turing and Shannon) identified games as an excellent test-bed for the application of Artificial Intelligence. For many years, research in this area focused on strategy games such as Chess, Checkers, and Go. Some notable AI successes (e.g. IBM’s Deep Blue and recently Schaeffer’s “perfect” draughts player) were hand-programmed, but more recently there has been greater emphasis on programs that learn for themselves, either by self-play temporal difference learning, or co-evolution. There is increasing interest in comparing these approaches, and in developing new hybrid techniques. Commercial games use increasingly sophisticated simulated environments, and present human players with world views approaching photo-realism. As virtual worlds become more complex, it becomes harder to design non-player characters (NPCs) with sufficiently interesting and believable behaviour. Current NPC control techniques mostly involve methods such as finite state machines, scripting, and search. Most of the behaviour is handprogrammed, involving much human effort to code and to test. Although games companies have occasionally used AI techniques such as genetic algorithms and neural networks, the methods used in games often lag behind the state of the art in machine learning. Similarly, machine learning and AI has much to benefit from games by applying and extending algorithms to more challenging environments than are usually encountered. Furthermore, as games utilise more realistic physics engines, there is scope for greater synergy between games and robotics research: the control of a humanoid robot, or of an NPC will share much in common.

One challenge the network will face is to find sufficient common ground to make collaboration attractive for both parties. This is because while academics may be most interested in how smart an NPC is, or how well it learns from its environment, a games company is in the business of selling games, and is therefore likely to be most interested in how intelligent NPC behaviour can be channelled into making games more enjoyable. Many academics are now eager to test their methods in the complex virtual worlds of modern console games, and – guided by industrial collaboration – this will bring great benefits both to the research community and to the games industry.