Modeling Primary and Secondary Emotions for a Believable Communication Agent Christian Becker, Ipke Wachsmuth Faculty of Technology University of Bielefeld 33594 Bielefeld Germany {cbecker,ipke}@techfak.uni-bielefeld.de Abstract. The integration of emotion and cognition in cognitive archi- tectures for embodied agents is a problem of increasing importance. In this paper, we describe how two separate modules for these tasks, as we employ them in our virtual human Max, can give rise to secondary emo- tions such as frustration and relief. The BDI-based cognitive module is responsible for appraisal as well as reappraisal of elicited emotions that our conversational agent Max becomes aware of. The emotion dynamics simulation system is driven by the valence information of every emotion and assures a general consistency of the simulated emotions over time by dynamically providing an awareness likelihood for every emotion. 1 Introduction and Motivation When trying to build socially intelligent agents, the integration of simulated emotions into an agent’s cognitive architecture seems to be unavoidable. In our work we follow the ideas of cognitive modeling with a cognitive architecture based on the BDI-theory and an emotion dynamics simulation system based on dimensional emotion theories. We further follow the ideas of [5] by distin- guishing “primary” and “secondary” emotions. “Primary” emotions are elicited as an immediate response to a stimulus, whereas “secondary” emotions are the product of cognitive processing. Up to now our set of simulated emotions has been limited to undirected, ”primary” emotions such as fear, sadness, anger and happiness. After we validated a desirable effect of these emotions empirically [4], we now want to let our agent control its own emotions to prepare him for social scenarios, where the application of coping strategies is necessary. In these situations the general limitation on “primary” emotions must also be overcome by further combining cognition and emotion to simulate “secondary” emotions such as frustration and relief as well. In the next section, we will start with an overview of related work along the lines of two major emotion research trends in psychology. In Section 3, we introduce our virtual agent’s modular architecture and describe our approach to simulate primary and secondary emotions as a combination of cognitive appraisal and an underlying emotion dynamics. We conclude this paper by discussing some advantages and possible drawbacks of our architecture.