Computational Models of Emotion

A computational model of emotion describes structures and processes related to emotion, mood, feeling, affect, or other related phenomena in sufficient detail such that it can be implemented on a computer. There are at least three kinds of computational emotion models: those that recognize emotion (e.g., recognize facial expressions), those that express emotion (e.g., generate predetermined facial expressions), and those that model human emotion processing (such as simulate how the inability to achieve a goal may lead to frustration). For models that simulate emotion, there are both high-level models that abstract away from details of the brain and attempt to describe all the interacting components involved in emotion, and low-level models that focus on the details of specific brain systems.

A model that recognizes emotion can be useful in improving human-computer interaction by providing additional information about a human user to the computer. For example, if a tutoring system can detect that a student is frustrated or confused, it can dynamically alter its teaching strategy to help the student. Sources of information that can be used to inform an emotion recognition system include facial expressions, vocal qualities, physiological measures (skin conductance, heart rate, blood pressure, etc.), and brain imaging.

A model that expresses emotion, usually through facial expressions or voice modulation, can improve interactions with humans by providing an indirect way of indicating the state of a computer system—whether it is being successful in its task or not and whether the human using the system is making progress on a task. Emotion expression also can enhance computational entertainment systems by making the behavior of virtual characters more realistic.

The remainder of this entry focuses on models described as simulating emotional processing. Rosalind Picard lists five criteria that must be met for a model to fully simulate emotion. In her words, such a system has emotion. In this article, we will finesse the debate over the differences between simulating and having emotion and concentrate on whether existing models achieve these criteria or not. Picard's five criteria for a model are as follows:

• 1.
  It engages in emotional behavior; some of its behavior is a consequence of its emotions.
• 2.
  It supports fast primary emotions, which are automatic emotional responses that can occur with little or no cognitive input.
• 3.
  It supports cognitively generated emotions, which are emotions that arise from more deliberate processing.
• 4.
  It has an emotional experience; it has subjective feelings.
• 5.
  There are body-mind interactions; emotions have an influence on physiological and cognitive processes and vice versa.

A system that achieves only a subset of these criteria can still be useful.

The criteria imply the need for not just a model of emotion but also a complete model of cognition. Indeed, most systems that claim to be nearly complete are implemented in cognitive architectures. Cognitive architectures are typically composed of a small number of interacting computational mechanisms (memories, processors, and interfaces) that give rise to complex behavior. Some architectures have been designed around emotion while others have been augmented with emotional mechanisms. Given the broad scope of behavior these systems [Page 166]attempt to model, cognitive architectures tend to be implemented at the symbolic, as opposed to neural, level (although hybrids exist).

...