Computer Science Department
School of Computer Science, Carnegie Mellon University


Guiding Interactive Drama

Peter Weyhrauch

January 1997

PhD. Thesis

Unavailable Electronically

Keywords: Computer science, artificial intelligence, interactive drama, interactive story, interactive fiction, interactive art, interactive entertainment, abstract adversary search, memoized search, sampled search, sampling search, SAS, SAS+, MFC, MMFC, aesthetic evaluation function, user experience, Oz Project

In traditional drama, an audience watches a story presented by characters on a stage. Interactive Drama changes the audience to a single User who enters that stage, interacting with the characters, and participating in the story.

This thesis concerns computer-based interactive drama. The computer system allows a User to interact with simulated worlds which are inhabited by dynamic and complex autonomous characters and shaped by a flexible, aesthetically pleasing story.

Like a traditional playwright, the artist who creates an interactive drama (ID) has a set of themes and ideas to be conveyed. Unlike the actions of characters in a play, however, the actions of the user in an ID are not known at story creation time, so an ID can not have a set script. The problem is to shape the experience of the User to conform to the set of themes and ideas to be conveyed, given that the actions of the User are not under the control of the artist. This work proposes solving this problem by dynamically monitoring and subtly guiding the experience of the User.

This dissertation describes an architecture called Moe that is designed to provide dramatic guidance. Moe uses abstract, adversary search with an aesthetic evaluation function to decide how and when to guide the User's experience. The first technical achievement of this work is a demonstration of the ability to capture a dramatic aesthetic (for one interactive drama) in an automated evaluation function that can examine an experience and determine its quality, much as a movie critic determines the quality of a film. This result is supported by statistically demonstrating a high degree of correlation (r = .87) between the evaluation function and the human artist.

The second main technical achievement of this work is the implementation of three strategies (SAS and SAS+, both based on random sampling; and MMFC, which uses memoization to implement a full-depth search) and a search state that seem capable of effectively modelling and guiding an interactive dramatic experience. The abstract search state includes the important aspects of the physical world, the characters, the presentation medium, and the User, including her mental, emotional, and physical states. Moe has not yet been connected to a running interactive experience, but instead Moe has been tested with a variety of Simulated user Types. On "average" users, SAS, SAS+, and MMFC are able to improve Users' experiences from the 50th percentile, to the 94th, 98th, and 99th percentile, respectively. This is a large and artistically meaningful increase.

This dissertation describes the first implementation of a system designed to provide centralized dramatic guidance in an interactive drama. Future work remains to determine whether the success of Moe in simulation can be effectively transferred to success with real Users.

210 pages

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by