 Springer-Verlag Heidelberg. Permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from Springer-Verlag Heidelberg. In B. Michaelis and G. Krell, editors, Pattern Recognition, Lecturenotes in Computer Science 2781, pages 148-155. Springer-Verlag, 2003. Learning Human-like Opponent Behavior for Interactive Computer Games Christian Bauckhage, Christian Thurau, and Gerhard Sagerer Technical Faculty, Bielefeld University, P.O. Box 100131, 33501 Bielefeld, Germany {cbauckha,cthurau,sagerer}@techfak.uni-bielefeld.de Abstract. Compared to their ancestors in the early 1970s, present day computer games are of incredible complexity and show magnificent graphical performance. However, in programming intelligent opponents, the game industry still applies techniques developed some 30 years ago. In this paper, we investigate whether opponent programming can be treated as a problem of behavior learning. To this end, we assume the behavior of game characters to be a function that maps the current game state onto a reaction. We will show that neural networks architec- tures are well suited to learn such functions and by means of a popular commer- cial game we demonstrate that agent behaviors can be learned from observation. 1 Context, Motivation, and Overview Modern computer games create complex and dynamic virtual worlds which offer nu- merous possibilities for interaction and are displayed using incredible computer graph- ics. Professional game development therefore has become expensive and time consum- ing and involves whole teams of programmers, authors and artists [2, 8]. However, de- spite all progress in appearance and complexity, when it comes to implementing intel- ligent virtual characters the game industry largely ignores scientific advances but still reverts to techniques known for more than 30 years [2]. Up to now, the most common techniques to control virtual characters are finite state machines (of admittedly complex topology) and scripts. From a player’s point of view this has two major drawbacks: (1) The actions of computer controlled characters often appear artificial since they just cycle through a fixed repertoire; this provokes repetitions and thus causes ennui and frustration. (2) Finite stated or scripted behaviors cannot generalize. Thus, if a human player acts unforeseen, i.e. interaction results in a game state not envisaged by the programmers, virtual characters tend to behave ’dumb’ [2]. Therefore –and certainly because of the popularity computer games enjoy among today’s students– creating intelligent opponents has attracted attention in AI research [3, 4, 11]. Especially Laird identifies a need for human-like behaving characters and heralds games as the ’killer application’ of artificial intelligence [8, 9]. And indeed, all cited contributions describe ontology based inference machines or reasoning mecha- nisms and thus are classical AI. In the following, however, we will argue that computer games also offer interesting problems and testbeds for the pattern recognition commu- nity. While the next section explains this claim in general, section 3 treats practical