Emotions as Durative Dynamic State for Action Selection

Emmanuel Tanguy, Philip J. Willis, Joanna BrysonIJCAI 2007

Introduction

In contemporary Artificial Intelligence, emotional modeling often relies on discrete, reactive triggers—a momentary "if-then" response to environmental stimuli. However, human emotions are inherently durative; they persist over time, accumulate, and fundamentally alter the context for future decision-making. This "lingering" effect means that an agent's reaction to an event is not just a function of the event itself, but of its internal emotional history.

In "Emotions as Durative Dynamic State for Action Selection," Tanguy et al. propose a framework to bridge the gap between reactive triggers and persistent emotional states. They introduce the Dynamic Emotion Representation (DER), a system designed to model how emotions build up, decay, and interact across multiple timescales.

The Dynamic Emotion Representation (DER) Framework

The core of the proposed architecture is a modular network where emotional dimensions are represented as units tracking intensity. To capture the complexity of human affect, the DER categorizes states into a three-tiered hierarchy:

  1. Behaviour Activations (Fast): Immediate, short-lived triggers such as a sudden smile or a flash of irritation.
  2. Emotions (Medium-term): States like happiness or anger that provide the immediate context for action selection and persist beyond the initial stimulus.
  3. Mood Variables (Long-term): Persistent states, such as tension or energy levels, that shift slowly and color the agent's overall disposition.

A critical innovation in the DER is the use of sigmoid functions as dynamic filters. These filters act as "gates" for emotional activation. For instance, a high "Happiness" state can shift the sigmoid threshold for "Anger," requiring a significantly stronger negative stimulus to trigger a hostile response. Furthermore, the system employs spreading activation, allowing emotions to mutually excite or inhibit one another (e.g., high tension amplifying anger while inhibiting happiness).

Experimental Methodology

The authors evaluated the DER using a scripted sequence of six emotional impulses: three positive (happiness) followed by three negative (anger). This sequence was processed through the DER under three distinct initial mood contexts:

  • Negative Mood: High tension, low energy.
  • Neutral Mood: Baseline levels.
  • Positive Mood: High energy, low tension.

To visualize the impact of these internal states, the DER was integrated with the Emotionally Expressive Facial Animation System (EE-FAS), a virtual "talking head" that translates emotional intensities into facial expressions.

Results and Discussion

The experiments demonstrated that the DER successfully produces context-dependent behavior. As shown in the comparative intensity graphs, the same set of stimuli resulted in vastly different emotional trajectories depending on the starting mood.

Emotional intensity profiles across Negative (a), Neutral (b), and Positive (c) mood contexts in response to the same stimuli (d).

In the Positive Mood condition, the agent maintained high happiness levels and effectively dampened the impact of the anger impulses. Conversely, in the Negative Mood condition, the agent showed minimal reaction to happy stimuli but experienced intense, lingering anger that built up with each subsequent negative impulse—a "last straw" effect.

The integration with EE-FAS further highlighted the system's realism. Rather than snapping between discrete expressions, the virtual agent displayed blended transitions and varying intensities of expression that reflected its underlying "emotional momentum."

Facial expressions generated by the EE-FAS system, illustrating how the same impulses result in different visual deliveries based on the underlying mood.

Conclusion

The DER framework provides a robust method for representing emotions as dynamic, durative states rather than simple reactions. By separating emotional appraisal from state representation and accounting for decay and interaction, the system allows for more nuanced and believable virtual agents. Future work aims to integrate these states more deeply into the goal-selection processes of autonomous agents, exploring how emotional "lingering" influences long-term survival and social interaction.

Summary updated

BibTeX
@inproceedings{tanguy2007emotions,
  author = {Emmanuel Tanguy and Philip J. Willis and Joanna Bryson},
  title = {{Emotions as Durative Dynamic State for Action Selection}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2007},
}
Figure