Emotional Agents need Formal Models of Emotion

Embedding a computational model of emotion in virtual agents is beneficial in a variety of domains. These domains include gaming, VR training, HCI and electronic tutors. Although these domains have different motives for embedding such a model, they share the same overall approach. Once the requirements for the agent are clear, first an emotion theory is chosen as basis for the computational model. Second, the model is implemented and embedded into the virtual agent. Candidate emotion theories are mostly cognitive, i.e. appraisal theories. Furthermore, theories used for this purpose are mostly structural descriptions -usually represented by text and tablesof the relations between events, evaluations of events and emotions. Structural descriptions are abstract, but computational models are concrete. In this paper we explain the nature of the gap between the level of abstraction of these structural descriptions of appraisal theories and computational models of emotion. We also show that this gap introduces several important problems that make it hard to evaluate the consistency between a computational model of emotion and the appraisal theory it is based on. Lastly, we propose a formalism to narrow this gap, which can be used to describe the structure of appraisal. We believe that our formalism stimulates the consistency of computational models based on appraisal theories and thereby increases the potential and plausibility of emotions in virtual agents and robots.