Continuous and Reinforcement Learning Methods for First-Person Shooter Games
Tony C. Smith, Jonathan Miles · PsycEXTRA Dataset · 2010
Machine learning is now widely studied as thebasis for artificial intelligence systems within computer games.Most existing work focuses on methods for learning staticexpert systems, typically emphasizing candidate selection. Thispaper extends this work by exploring the use of continuous andreinforcement learning techniques to develop fully-adaptivegame AI for first-person shooter bots. We begin by outlining aframework for learning static control models for tanks withinthe game BZFlag, then extend that framework using continuouslearning techniques that allow computer controlled tanks to adaptto the game style of other players, extending overall playability bythwarting attempts to infer the underlying AI. We further showhow reinforcement learning can be used to create bots that learnhow to play based solely through trial and error, providing gameengineers with a practical means to produce large numbers ofbots, each with individual intelligences and unique behaviours;all from a single initial AI model.