A. Bryan Loyall

Believable Agents: Building Interactive Personalities Degree Type: Ph.D. in Computer Science
Advisor(s): Joseph Bates
Graduated: May 1997

Abstract:

In the traditional character-based arts of film, books, animation, or theater, the audience is presented with vivid, powerful, personality-rich characters such as Rick from @i(Casablanca) or the Genie or Flying Carpet from Walt Disney's @i(Aladdin). Unless one can afford a troop of improvisational actors (and not even then for characters such as the Flying Carpet), one is only able to watch these rich characters. @i(Believable agents) allow people to not just watch, but also interact with such powerful, personality-rich characters.

Believable agents are a combination of autonomous agents and believable characters from the traditional character-based arts, such as animation, film or literature. They are accurately described both as autonomous agents with the same powerful properties as characters from the arts, and as computer-based, interactive versions of personality-rich characters. Believable agents could be powerfully used for art, entertainment, as part of interactive story systems, social user interfaces, or teaching systems.

This dissertation presents my progress toward the creation of believable agents. The first result of this dissertation is a study of the problem to be solved. This study has two parts: a study of believability in traditional character-based arts, and a study of believability for agents, primarily drawing on experience in multiple attempts to construct such agents.

The second result of the dissertation is an architecture called Hap specifically designed to support the requirements and expression of believable agents. This support includes a control of real-time interactive animation, and architectural support for many of the requirements of believable agents. The architecture is also a unified architecture for the agent in that all aspects of an agent's mind are expressed directly in Hap. By doing this the other aspects, such as arbitrary computation, computation of emotions, and recognizing patterns of activity over time for higher-level sensing, all inherit Hap's properties for believability.

The third result of this thesis is an approach for natural language generation for believable agents. This result includes an analysis of the requirements for natural language in believable agents, an architecture for supporting these requirements, and an example grammar and behaviors that use this architecture as part of an agent's behavior.

A number of agents have been built using this framework. One is described in detail to show how the technologies described in the dissertation can be used together to build complete believable agents. A public exhibit of some of the constructed agents provides evidence that this approach can achieve a certain level of believability, and that others can learn the framework and use it to build believable agents.

Keywords:
Artificial intelligence, believable agents, interactive characters, personality, active language, embodied language and action, situated natural language generation, art, entertainment, Oz