Abstract
This paper presents the "Affective Reasoner," a symbolic AI framework designed to simulate personality and emotion in autonomous agents. By employing a "shallow" model based on cognitive appraisal theory rather than biological simulation, the authors demonstrate that believable social interactions can be computationally modeled. The system maps situational appraisals to specific emotional states and multi-modal expressions, enabling agents to engage in complex social behaviors such as empathy and strategic bluffing.
Introduction
The quest for "believable agents" in artificial intelligence often hits a wall when it comes to social intelligence. Traditional AI systems of the early 1990s were proficient in logic but lacked the "affective state" necessary for human-like interaction. This paper, originally presented at AAAI-94, explores a pragmatic approach to this problem: the Affective Reasoner (AR). Instead of attempting to replicate the biological complexity of the human brain, the AR uses a shallow symbolic model to map situations to emotional responses.
The Problem of "Cold" AI
The authors argue that computers typically lack "concerns" or goals that drive behavior in a way users find relatable. Without an understanding of how a user feels or a set of internal principles, AI remains socially inept. The challenge addressed here is creating a model that is both computationally tractable for contemporary hardware and sophisticated enough to foster engagement. The goal is not machine consciousness, but functional believability.
Methodology: The Affective Reasoner (AR)
The AR platform is built upon the Ortony, Clore, and Collins (OCC) model of cognitive appraisal. This theory posits that emotions are not random but are the result of how an agent interprets—or "appraises"—an event relative to its goals and principles.
As shown in Table 1, the system categorizes emotions into various types based on specifications like the appraisal of a situation as an event or an accountable act. For instance, "joy" is defined as being pleased about an event, while "reproach" involves disapproving of another's act. Agents interpret their environment through "appraisal frames" that represent their unique personality traits.
To manifest these internal states, the AR utilizes multi-modal expression:
- Schematic Faces: A library of ~70 faces that can morph into over 3,000 variations.
- Inflected Speech: Text-to-speech that adjusts speed, pitch, and volume based on the current emotion.
- Music Indexing: A MIDI-based system that selects music matching the agent's mood.
Key Insights and Results
A central thesis of the work is that "shallow" modeling is a feature, not a bug. By treating emotion as a byproduct of goal-driven behavior, the system can simulate complex social dynamics. For example, friendship is modeled through shared goals: an agent feels "happy-for" a friend's success and "pity" for their failure.
Empirical tests of the system yielded promising results:
- Speech Recognition: The system correctly detected 188 out of 198 emotion words in context. When users spoke with theatrical inflection, the system achieved nearly 100% accuracy in recognizing seven broad emotion categories.
- Social Simulation: In a poker domain, agents successfully used emotional states to bluff (leading to "fear" or "hope"), gloat after a win, or feel "remorse" after a loss, creating a more engaging gaming experience.
- Efficiency: The entire multi-modal system—including facial morphing and speech recognition—was capable of running concurrently on a standard 66 MHz PC.
Conclusion and Future Directions
While the model successfully creates believable interactions, it faces limitations in its deep understanding of human experience; it recognizes the word "sad" without truly understanding the biological experience behind it. Furthermore, the speech recognition required a degree of "theatricality" from users to be fully effective. Future work points toward applying these affective models to military stress training, tutoring systems, and automated storytelling, where emotional themes can serve as organizing principles for narrative plots.
Summary updated