Affective computing has proven to be a viable field of research comprised of a large number of multidisciplinary researchers, resulting in work that is widely published. The majority of this work consists of emotion recognition technology, computational modeling of causal factors of emotion and emotion expression in virtual characters and robots. A smaller part is concerned with modeling the effects of emotion on cognition and behavior, formal modeling of cognitive appraisal theory and models of emergent emotions. Part of the motivation for affective computing as a field is to better understand emotion through computational modeling. In psychology, a critical and neglected aspect of having emotions is the experience of emotion: what does the content of an emotional episode look like, how does this content change over time and when do we call the episode emotional. Few modeling efforts in affective computing have these topics as a primary focus. The launch of a journal on synthetic emotions should motivate research initiatives in this direction, and this research should have a measurable impact on emotion research in psychology. In this paper, I show that a good way to do so is to investigate the psychological core of what an emotion is: an experience. I present ideas on how computational modeling of emotion can help to better understand the experience of emotion, and provide evidence that several computational models of emotion already address the issue.
TopIntroduction
In this position paper I argue that efforts in computational modeling of emotion should focus more on returning results to psychology, and I propose a research direction that can achieve this.
Computational models of natural phenomena are useful for two main reasons. First, the model itself can be used to simulate and predict the phenomenon that is modeled. Consider for example the weather. A detailed computational model of clouds, temperature fronts, pressure systems and geological factors predicts the weather for the next couple of days. This is obviously useful, and it is a quality of the model related to its usefulness when applied in a particular context (predicting the weather). Second, any model is the instantiation of a theory or set of hypotheses, whether these hypotheses are simple or complex, widely validated or new. For a model to be computational, it needs to be executable by a computer (the model must be a computer program that can run). Regardless of the particular peculiarities of the computer system that is used to run the program, a fairly detailed description of the model is always needed. The computer needs detailed step-by-step instructions that match the model. Therefore, a computational model is a detailed instantiation of a theory or set of hypotheses. The predictions produced by a running computational model are predictions of the theory that a model is based on. Obviously, the usefulness of these predictions critically depends on two factors: the credibility of the theory used as basis for the model, and the credibility of the extra assumptions that were needed to build the computational instantiation of that model (the correctness of the implementation). This means that a computational model also has an intrinsic quality: it can be used to evaluate a theory. For example, predictions produced by a computational model of the weather can be used to evaluate the theory of the weather that underlies the computational model. Incorrect predictions motivate changes to the theory.
This view of computational models is not different in the area of affective computing. Affective computing is a research field that is concerned with the development and use of computational models of emotion. Typically, such models are used in the domains of emotion recognition, emotion elicitation (production), and emotion effects (e.g., on cognition, behavior). In following sections I will discuss emotion and affective computing in more detail, but first I will give a more concrete example of a computational model of emotion elicitation that has both qualities, the applied quality any model has that produces useful output, and the intrinsic quality a model has provided that it is grounded in theory. Consider the work by (Gratch & Marsella, 2001). They propose a model of emotion elicitation, implemented in a pedagogical software agent. The role of the software agent is to guide a trainee through a virtual-reality based training session. The software agent can—partly as a result of the model of emotion—deliver a more believable training simulation (more believable than training without emotional agents). This hopefully results in a training session that better matches reality, giving participants in the training a better preparation for real life situations. This is an applied quality of the computational model of emotion. If the modelers did a faithful job of transforming the theory they used as basis for their model into a computational instantiation then the behaviors of the agent are in essence predictions of the theory underlying the computational model. This is the intrinsic quality of the computational model of emotion. Predictions of this kind can be used as a motivation for changes to the emotion theory underlying the model (Broekens, DeGroot, & Kosters, 2008).