R. Shumaker (Ed.): Virtual and Mixed Reality, Part II, HCII 2011, LNCS 6774, pp. 289–296, 2011. © Springer-Verlag Berlin Heidelberg 2011 Why Can’t a Virtual Character Be More Like a Human: A Mixed-Initiative Approach to Believable Agents Jichen Zhu 1 , J. Michael Moshell 1 , Santiago Ontañón 2 , Elena Erbiceanu 1 , and Charles E. Hughes 1 1 University of Central Florida, USA 2 Artificial Intelligence Research Institute (IIIA-CSIC), Spain jzh@mail.ucf.edu, jm.moshell@cs.ucf.edu, santi.ontanon@gmail.com, eelena@eecs.ucf.edu, ceh@cs.ucf.edu Abstract. Believable agents have applications in a wide range of human computer interaction-related domains, such as education, training, arts and entertainment. Autonomous characters that behave in a believable manner have the potential to maintain human users’ suspense of disbelief and fully engage them in the experience. However, how to construct believable agents, especially in a generalizable and cost effective way, is still an open problem. This paper compares the two common approaches for constructing believable agents — human-driven and artificial intelligence-driven interactive characters — and proposes a mixed-initiative approach in the domain of interactive training systems. Our goal is to provide the user with engaging and effective educational experiences through their interaction with our system. Keywords: Mixed-initiative system, character believability, interactive storytelling, artificial intelligence, interactive virtual environment. 1 Introduction Interactive computer characters are becoming widely used in digital systems in a range of domains. Some applications include, for example, virtual patients used in training medical students’ communication skills [1], anthropomorphic robotic assistants that can support manual tasks in industrial environments [2], and non-player characters (NPC) in computer games that provide players with dramatic experiences [3]. These virtual or robotic characters are an important part of human centered computing not merely because they provide users with a set of familiar metaphors and conventions to interact with new technology. Like Disney’s animated characters [4], they also have the powerful potential to engage users in emotional and personal ways. With a significant improvement of their appearance thanks to recent development of real-time computer graphics technology, there is an increasing demand for these characters to behave in ways that are believable to the users. However, how to construct believable agents, especially in a generalizable and cost effective way, is still an open problem. This paper compares the two common approaches