This paper briefly proposes a theory of emotions that clarifies their role in architectures for intelligent agents. The account posits that emotions take the form of symbolic cognitive structures, that generic emotional rules produce concrete instances of these relational concepts, that such rules underlie both generation and understanding of emotions, and that their results play a metacognitive role in influencing behavior. In this framework, emotions are central to high-level, human-like cognition.
∙ ∙ ∙ ∙
Afect . The valence and intensity for some experience.
Mood. A global variant of afect for the entire cognitive system.
Emotion. A relation among goals and beliefs for an event or object.
Feeling. An afective or hormonal response associated with an emotion.
Elsewhere we discuss the need to incorporate emotions into cognitive architectures
One key question is whether emotions are better explained at the architectural level, and thus require specialized structures and procedures, or at the knowledge level, and thus require only the addition of content encoded in established representations. Another concerns what functions emotions can serve in intelligent agents, and thus what capabilities they support. In the sections that follow, we propose some tentative answers to these questions.
Before we can discuss processes that produce emotions, we must first consider how to represent
them.
Most dimensional models characterize emotions as points in a three-dimensional space:
pleasure (measure of valence); arousal (level of afective activation); and dominance (measure
of control). Synthetic characters often use such “PAD” models
Appraisal models view emotions as inferred relations among mental structures based on
situations.
Our framework maps such “appraisal frames” onto concepts that reside in the PUG
architectures’s long-term memory
As an example, consider the familiar emotional concept disappointed, which relates a person P to some situation or event S. An instance of this emotion can arise when P wants S to become true, P expected S to take place, and yet P believes that it did not actually occur: ∙ disappointed(P, R) :- wants(P, R), expected(P, R), believes(P not(R)) .
Similarly, we say that a person P is jealous of another person Q if P wants to possess an object S, believes that he does not possess S, and believes that Q does possess it. Complex emotions are specializations of basic ones that involve more conditions.
Although emotions have a clear structural element, their instances have an associated afective score that includes a valence and intensity, with the latter changing over time. Similarly, emotional concepts incorporate a value function that specifies this afect as a function of matched elements. Most appraisal theories include these features, but they do not map onto aspects of a cognitive architecture like PUG that includes them for other reasons.
Our theory also takes positions about how emotional structures are processed, which we can separate into two high-level cognitive tasks. One is generation, which produces emotions for the primary agent, such disappointed(John, failed(John, CompSci101)). The other is understanding, which infers the emotions of other agents, such as belief(John, disappointed(Jane, failed(John, CompSci101))). This maps onto the classic distinction between plan generation and plan understanding. Both appear necessary for a full account of emotion’s relation to cognition.
We maintain that a single PUG mechanism, conceptual inference, underlies both generation of the primary agent’s emotions and inference about those of others. This uses a variety of relational pattern matching like that in logic programming, although an alternative would use abduction rather than deduction. This draws on processes available in the existing architecture, which ofers further evidence that emotions are knowledge-level phenomena.
However, this raises a question about why afective scores are more intense for our own emotions than those inferred for others. Presumably, this diference results from coeficients in the value function associated with each emotional concept, which will be higher for the primary agent and lower for inferences about other agents. This may also explain why memories of past emotions are typically less intense than the original experience. This suggests that the distinction between ‘hot’ and ‘cold’ emotions is not a dichotomy but rather a continuum.
A full theory should clarify not only how the architecture generates and infers emotions, but also how they influence other cognitive processing. Thus, it should specify how emotional instances impact the agent’s physical behavior (e.g., crying about loss or punching someone) or its cognitive processing (e.g., changing goal priorities or invoking planning). This supports the view that emotions are not evolutionary relics, but rather high-level regulators of cognition.
We postulate that this relates to people’s ability to think about thinking – metacognition.
Recall that emotional concepts specify abstract relations among goals, beliefs, and expectations,
which are matched during conceptual inference. This suggests a promising hypothesis:
∙ Emotions play a metacognitive role that operates over and influences base-level cognition .
That is, emotional processing inspects traces of basic cognition and alters its course in response.
This view follows
The PUG architecture already incorporates long-term structures called motives that match
against domain-level beliefs, generate top-level goals, and compute the latters’ values. Elsewhere,
we have proposed motives that match against emotions and generate goals which may involve
changing emotions of the primary agent or others
Our theory of cognition and emotion draws on ideas that have appeared elsewhere in the
literature, many of them discussed by Marsella, Gratch, and Petta (2010) and by
One notable diference is that their accounts assume a fixed set of appraisal frames, each
associated with a distinct emotion, whereas our theory allows an arbitrary number of
emotional concepts, some defined in terms of others. In this sense, it comes closer to
In summary, our computational account of emotion brings together ideas from appraisal theory and cognitive architectures. The most important claims of the framework, which it shares with some other work that follows similar lines, are that: ∙ Emotions are symbolic cognitive structures with numeric annotations; ∙ Generic emotional rules generate specific instances of these concepts; ∙ These rules are used to generate emotions and to infer those of others; ∙ Emotions play a metacognitive role in influencing cognition and behavior; ∙ These influences are mediated by generic motives that assign values to goals. In this theory, emotions are not irrational vestiges of evolution. Rather, they are linked directly to structures and processes in the cognitive architecture, specifically ones in the PUG framework, which should require only minor extensions to incorporate them.
We maintain that PUG’s division of knowledge into concepts, skills, and motives makes it especially suitable for explaining emotions and their relation to both cognition and personality. However, we must still demonstrate the architecture’s ability to reproduce emotion-related phenomena in realistic scenarios that are similar to ones humans encounter. We must also work out details about how to calculate emotional intensities as a function of time, diferent agents, and other factors, but we remain optimistic about the framework’s potential in this arena. The research reported here was supported by Grant No. FA9550-20-1-0130 from the US Air Force Ofice of Scientific Research, which is not responsible for its contents. We thank John Laird and the reviewers for constructive comments that improved the paper.