A Review of Agent Emotion Architectures

This review, published in the journal Eludamos in 2008, surveys computational models of emotion for software agents in games and simulations. Its central argument is that agent emotional AI has advanced more slowly than graphics, sound, and control technologies, and that a key obstacle is the lack of accessible, psychologically grounded emotion architectures for game developers who lack a background in psychology. The authors position the work as a foundation for developers who wish to draw on academic AI research for the next generation of games.

The review examines three prominent architectures repeatedly cited in the literature: CogAff from the University of Birmingham, a three-layer reactive, deliberative, and meta-management model that maps primary, secondary, and tertiary emotions to its layers; the Tok project from Carnegie Mellon, which pairs the HAP reactive architecture with the Em emotion module, itself a scaled-down version of the OCC model, and which includes emotional decay over time; and Émile, from the University of Southern California, which combines appraisal and coping mechanisms and lets agents observe the emotional states of others. The authors also note related work such as Bartneck's analysis of the OCC model, which proposes a five-stage treatment covering classification, quantification, interaction, mapping, and expression, and identifies gaps including the absence of a history function and emotion interaction. Across these projects, the review finds that the emotional components are tightly integrated with agent decision-making, making them difficult to separate from the broader agent system.

The authors conclude that commercial game engines already provide behaviour, sensing, and planning, so developers most need either a middleware emotion module compatible with those engines or an architecture grounded in sound psychology yet written for non-specialists. They also flag terminological confusion, such as the blending of personality labels with emotions, and the ambiguity of terms like distress and hope. The review's evidence base is a literature survey rather than an empirical study, and its authors observe that most surveyed systems were demonstrated at small scale and few were tested in real-time commercial settings.