A Logic of Emotions for Intelligent Agents

Bas R. Steunebrink, Mehdi Dastani, John-Jules Ch. MeyerAAAI 2007 · OpenReview

Abstract

Traditional artificial intelligence models often struggle with the "excessive deliberation" problem, where rational agents become paralyzed by logical possibilities. This paper proposes a formal bridge between psychology and computer science by integrating the OCC (Ortony, Clore, & Collins) model of emotions into the KARO logical framework. By defining emotions such as hope and fear as mathematical fluents, the authors demonstrate how these states serve as vital heuristics for agent decision-making, specifically in determining when to persist with a plan and when to initiate replanning.

Introduction

In the design of autonomous agents, the Belief-Desire-Intention (BDI) framework has long been the standard for modeling rationality. However, purely rational agents often lack the efficiency of human decision-makers. As neuroscientists like Antonio Damasio have observed, emotions are not merely "noise" in human cognition; they are essential shortcuts that prevent over-thinking.

This research addresses the challenge of formalizing these "vague" emotional concepts into rigorous mathematical definitions. By treating emotions as logical fluents, the authors provide a framework where an agent's internal state—its "feelings"—directly informs its computational efficiency and behavioral persistence.

The Formal Framework: Extending KARO

To ground emotions in logic, the authors utilize the KARO framework, which synthesizes several specialized logics:

  • Dynamic Logic: Used to model actions and their subsequent effects on the environment.
  • Epistemic and Doxastic Logic: Used to represent the agent’s knowledge and beliefs, respectively.
  • BDI Operators: Used to track the agent's goals (desires) and its commitments to specific actions (intentions).

The core contribution of this work is the extension of KARO to include emotional fluents. The authors map the 22 emotions defined in the OCC model—a psychological hierarchy that classifies emotions based on reactions to objects, actions, and events—into this formal language.

Case Study: Hope and Fear

While the framework covers the full OCC hierarchy, the paper focuses on the interplay between hope and fear as a primary case study for agent behavior. These are defined through logical bi-implications:

  • Hope: An agent is said to hope for a goal if it possesses a plan, intends to execute that plan, and remains committed to the outcome.
  • Fear: An agent experiences fear regarding a failure if it maintains hope for a goal but simultaneously believes that the current plan or environment might lead to the goal’s negation.

Key Insights and Logical Results

Emotions as Goal-Tracking Mechanisms

In this model, emotions are not transient moods but functional evaluations of progress. Hope represents the logical "pleasure" of possessing a viable plan, while fear represents the "displeasure" of identifying potential failure points. This formalization leads to several significant insights:

  1. The Determinism Paradox: The authors prove that in a perfectly deterministic world where an agent has perfect information, the agent will never experience fear. Fear is shown to exist exclusively in the gap between an agent's intentions and the uncertainty of the outcome.
  2. The Hope-Fear Complement: The logic dictates that fear is dependent on hope. An agent cannot fear the failure of a goal it does not first hope to achieve.
  3. Persistence and Rationality: The authors demonstrate that hope persists only as long as the plan (the "prospect") is active. Furthermore, they prove that a rational agent with consistent beliefs will not hope for the impossible or fear the inevitable.

Practical Application: Emotions as Replanning Heuristics

The paper proposes that these logical definitions solve the problem of non-determinism in agent implementation. Instead of an agent randomly selecting actions when faced with uncertainty, it can use the relative intensity of its emotions to govern its control loop:

  • If Hope > Fear: The agent continues executing the current plan.
  • If Fear > Hope: The agent halts execution and invokes a replan function.

This turns emotional states into a formal deliberation rule, effectively using "feelings" as a trigger for meta-cognitive shifts in strategy.

Limitations and Future Directions

Despite the success in qualitative formalization, the authors identify several areas for further development:

  • Quantitative Modeling: While the logic defines when an emotion occurs, it does not yet handle the intensity (the "how much") of the emotion. Moving from a qualitative to a quantitative model is necessary to fully implement the replanning heuristic.
  • Social Complexity: The current model focuses on a single agent. Future work is required to formalize social emotions like anger, gratitude, or pity, which involve interactions and beliefs about other agents.
  • Physical Implementation: The authors intend to apply this logic to the iCat social robot. By embedding these logical emotions into a physical platform, they aim to create more "believable" and efficient interactions in practical scenarios such as elder care or domestic assistance.

Summary updated

Abstract

This paper formalizes a well-known psychological model of emotions in an agent specification language. This is done by introducing a logical language and its semantics that are used to specify an agent model in terms of mental attitudes including emotions. We show that our formalization renders a number of intuitive and plausible properties of emotions. We also show how this formalization can be used to specify the effect of emotions on an agent's decision making process. Ultimately, the emotions in this model function as heuristics as they constrain an agent's model.
BibTeX
@inproceedings{steunebrink2007logic,
  author = {Bas R. Steunebrink and Mehdi Dastani and John-Jules Ch. Meyer},
  title = {{A Logic of Emotions for Intelligent Agents}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2007},
  url = {https://openreview.net/forum?id=H14T5Ax_-S},
}
Figure