Building Emotional Characters for Interactive Drama

W. Scott ReillyAAAI 1994

Introduction

In the development of interactive drama, the creation of believable virtual agents is a primary challenge. Characters that lack emotional depth often appear "flat" or "wooden," breaking the immersion for the user. The Oz project addresses this by developing architectures that prioritize artistic believability over strict psychological realism. This paper introduces Em, an emotion subsystem integrated into the Tok agent architecture, designed to provide virtual characters with idiosyncratic emotional lives and expressive behaviors.

The Tok/Em Architecture

The system is divided into two primary components: Tok, which handles high-level goals and action selection, and Em, which manages the agent's internal emotional state. This separation allows the agent to maintain a coherent set of objectives while its internal state fluctuates based on environmental stimuli.

Emotion Generation

Em utilizes the OCC model (Ortony, Clore, & Collins, 1988) to generate emotions based on the agent's cognitive appraisal of its environment. When an agent's goal is achieved or thwarted, or when social standards are met or violated, Em generates specific emotional states such as joy, distress, pride, or shame. This allows the agent to react dynamically to the unfolding narrative in a way that reflects its personal stakes.

Behavioral Features and Idiosyncrasy

A core innovation of this research is the introduction of Behavioral Features (BFs). Rather than having an emotion directly trigger a specific action, emotions are mapped to BFs. This "middle layer" allows for character-specific responses—a concept known as idiosyncrasy.

For example, while two characters might both experience "anger," their behavioral manifestations can differ significantly based on their defined personality:

  • A "bully" character might map anger to an aggressive BF, leading to confrontational actions.
  • A "shy" character might map the same anger to a withdrawn BF, leading to avoidance or silence.

These BFs influence the Tok architecture through several mechanisms:

  1. Goal Creation: An angry agent might gain a new goal to retaliate.
  2. Plan Selection: A sad agent might choose a slower, more deliberate path to a destination.
  3. Priority Management: Fear might elevate "self-preservation" goals above all others.
  4. Action Style: Emotion changes how an action is performed. The distinction between a standard "walk" and a "stomp" is a powerful tool for non-verbal communication of internal states.

Implementation and Results

The architecture was validated through several experimental agents, moving from simple proof-of-concepts to more complex social scenarios:

  • Lyotard: A virtual cat that served as the initial demonstration of the Em/Tok integration, showing that the system could produce reasonably believable animal-like behavior.
  • The Woggles: Three ellipsoidal characters in a real-time graphical environment. Despite their simple geometric forms, informal evidence from hundreds of users suggested the Woggles were perceived as engaging and believable. This proved that complex graphics are not a prerequisite for believability if the emotional logic is sound.
  • The Hold-up Story: A scenario involving a gunman and a cashier, designed to test the limits of the system in high-stakes, nuanced human interactions.

Discussion and Future Work

The Em subsystem demonstrates that a rule-based approach to emotion can successfully create the "illusion of life" in virtual agents. By focusing on action style and idiosyncratic mappings, the researchers provided a framework that is accessible to artists and flexible enough for diverse character types.

However, limitations remain. In the current implementation, many of the mappings between Behavioral Features and specific goals were hand-coded, which limits scalability. Furthermore, while qualitative feedback was positive, the field still lacks rigorous, standardized metrics for measuring "believability." Future work aims to automate these behavioral mappings and refine the evaluation process for interactive storytelling.

Summary updated

BibTeX
@inproceedings{reilly1994building,
  author = {W. Scott Reilly},
  title = {{Building Emotional Characters for Interactive Drama}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure