Scalable and Flexible Appraisal Models for Virtual Agents

Computational models of emotion are useful in a variety of domains, including games, virtual realty training and HCI to name a few. Many of these models are inspired by appraisal theory. Most appraisal theories share with virtual agents the assumption that beliefs, desires and intentions are the basis of reasoning and thus of the emotional evaluation of the agent's situation. Consequently most computational models of emotion are deeply embedded into the agent model. In this paper we address the problem of how to emotionally instrument a system in a modular and extensible way, so that emotional sophistication can be added incrementally to a system. We propose a solution based on a modular, signalbased approach to computational emotions that allows us to develop scalable appraisal models that are easily added to non-emotional systems. Our approach allows runtime tradeoff between emotional quality and performance, which makes it particularly useful in domains in which available computation time is unknown, like the gaming domain. We present experimental results that back-up our approach.