How to Determine the Utility of Emotions

In this paper, we describe a new methodology for determining the utility of emotions. After briefly reviewing the status quo of emotional agents in AI, we describe the methodology and demonstrate it by showing the utility of “anger” for biologically plausible foraging agents in an evolutionary setting. Background on Emotions and AI Evidence from psychology (Frijda 1986; Izard 1991; Scherer, Schorr, & Johnstone 2001), neuroscience (Damasio 1994; LeDoux & Fellous 1995; Panksepp 2000; Hamm, Schupp, & Weike 2003), and ethology (Lorenz & Leyhausen 1973; McFarland 1981) suggests that emotions play several crucial roles in biological organisms. Especially in humans, they seem to be deeply intertwined with cognitive processing (e.g., they can bias problem solving strategies in humans (Bless, Schwarz, & Wieland 1996; Schwarz ) or help to evaluate a situation quickly (Kahneman, Wakker, & Sarin 1997; Damasio 1994; Clore, Gasper, & Conway 2001)). Finally, and most importantly, emotions are crucially involved in social control (Frijda 2000; Cosmides & Tooby 2000) ranging from signaling emotional states (e.g., pain) through facial expressions and gestures (Ekman 1993) to perceptions of emotional states that cause approval or disapproval of one’s own or another agents’ actions (relative to given norms), which can then trigger corrective responses (e.g., guilt). Yet, there is not even agreement among emotion researchers about how to construe basic emotions or whether the concept is coherent (Ortony & Turner 1990; Griffiths 1997). The difficulties with emotion concepts are also reflected in AI, where different forms of emotions have been been investigated to varying degrees ever since its beginnings (e.g., (Toda 1962; Simon 1967; Pfeifer & Nicholas 1982; Dyer 1987; Pfeifer 1988)). Over the recent years, various “believable synthetic characters and life-like animated agents” (e.g., (Bates 1994; Hayes-Roth 1995; Maes 1995; Lester & Stone 1997; Rizzo et al. 1997)), “emotional pedagogic agents” (e.g., (Gratch 2000; Shaw, Johnson, & Ganeshan 1999; Lester et al. 1997; Okonkwo & J.Vassileva 2001; Conati 2002)), “emotional virtual agents and robots” (e.g., Copyright c © 2004, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. (Bates, Loyall, & Reilly 1991; Velásquez 1999; Michaud & Audet 2001; Breazeal 2002; Arkin et al. 2003)), and “computational models of human emotion” (e.g., (Eliott 1992; Cãnamero 1997; Wright 1997; Allen 2001; Marsella & Gratch 2002)) have been proposed.1 Yet, there are divergent views among all these researchers about what it means to implement emotion in agents (e.g., (Ventura & Pinto-Ferreira 1999; Wehrle 1998; Picard 2001; Scheutz 2002a)). Most of this work in AI has focused on what could be called effect models of emotion. Effect models implement only overt, observable effects of emotional behavior. They are intended to get the “input-output mapping” of a given behavioral description right. In the extreme case, such a mapping could be as simple as that employed in an animated shopping agent which displays a surprised face if the user attempts to delete an item from the shopping basket. Many architectures of so-called “believable agents” (e.g., (Hayes-Roth 1995; Scheutz & Römmer 2001; Rizzo et al. 1997; Loyall & Bates 1997) for simulated agents and (Shibata & Irie 1997; Breazeal 1998; Velásquez 1999; Michaud & Audet 2001; Murphy et al. forthcoming) for robots) are part of this group, where the primary goal is to induce the belief in the human observer that the agent is in a particular emotional state. The main problem with effect models is that they are silent about the role of emotion in agent architectures. They may or may not actually implement emotional processes to achieve the desired overt behaviors. And if they do, the implemented states are often labeled with familiar terms, without specifying how the implemented states differ from those usually denoted with these terms (McDermott 1981; Scheutz 2002a). A state labeled “surprise”, for example, may have very little in common with the complex processes underlying notions of “surprise” in humans and various animals (i.e., the violation of a predicted outcome (Ortony, Clore, & Collins 1988; Macedo & Cardoso 2001)), if it is functionally defined to be triggered by loud noises (Velásquez 1997a; 1997b) (for such a state, “startle” would be the more appropriate label). Effect models are, therefore, inadequate for determining the utility of emotions in agent architectures. This list is only a brief excerpt of the recent literature and by far not complete, see also (Trappl, Petta, & Payr 2001; Hatano, Okada, & Tanabe 2000; Pfeifer 1988). Process models of emotion, on the other hand, are applicable as they are intended to model and simulate aspects of emotional processes (typically in humans) as they unfold (Peschl & Scheutz 2001), following predictions of psychological or neurological theories of emotion (Scherer 1993; Panksepp 1998; Ortony, Clore, & Collins 1988)). Process models are much more complex than effect models, given that they focus on the internal processes of an agent’s control system, and are typically only implemented in simulated agents (e.g., (Wright 1997; Marsella & Gratch 2002; Cãnamero 1997; Allen 2001; McCauley & Franklin 1998)). The problem with current process models is threefold: for one, they do not provide or use a conceptual framework to characterize the implemented emotional states (i.e., what kind of state it is they implemented and what it takes in general to implement such a state), nor do they investigate variations of such states sytematically. And most importantly, they typically do not evaluate emotional states with respect to their utility (neither by varying architectural parameters nor by comparing them to other implementations of the same task). Hence, the potential of the states, other than to be present in a particular model, remains unclear. Evaluating the Utility of Emotional