Grants and Contracts Details
Project Summary Virtual environments increasingly employ interactive narratives for education, training, therapy, and entertainment. They must balance the player’s agency, the ability to make meaningful choices, with the author’s structure, design constraints that guide the player toward an intended learning experience. That balance has typically been achieved by human authors investing exponential effort in branching stories at design time or time-consuming manual effort at run time. An intelligent author agent capable of understanding its audience will lay the foundation for a new generation of virtual environments that provide agency while maintaining structure, allowing human players to share authorship with arti?cial authors and engage with interactive narratives that are personal, structured, and effective. As this next phase of my research career begins, I propose a computational model of how the participants in an interactive narrative ask questions and form answers as they reach a mutual understanding that allows them to cooperate. The model is based on the explicit question answering methods psychologists use to measure understanding after a narrative, except that it happens automatically and implicitly as the narrative is occurring and is applied in both directions-from player to author and author to player. The ability of humans and agents to perform this Mutual Implicit Question Answering (MIQA) task is a quantitative measure of their ability to provide agency, maintain structure, and share authorship that can be applied to the many existing interactive authorship techniques and the new ones produced by this research. Intellectual Merit This research represents a transformative shift in both the PI’s career and the ?eld of computational models of narrative toward virtual agents that are not just understood by their audience but also understand their audience. Shared authorship has long been a stated goal for this ?eld, but is rarely studied directly or measured quantitatively. This research de?nes a domain-independent task that measures shared authorship in human/human and human/computer narrative interactions that will unify efforts in designing and evaluating interactive narrative authorship. The key that allows an agent to perform Mutual Implicit Question Answering automatically is the merger of two previously separate research efforts in computational models of narrative: fast generative narrative planning algorithms that reason about the future and analytical models of question answering and memory that reason about the past. Together they form a domain-independent model that makes authorship decisions in real time based on what can happen next, what each participant expects, and how new actions affect those expectations. Broader Impacts This research enhances the infrastructure of scienti?c education by training a doctoral research assistant and two undergraduate research assistants at the most economically and ethnically diverse public university in Louisiana in a highly interdisciplinary ?eld that attracts students to AI research. Students in two annual game design courses will iteratively design, develop, and test an interactive narrative virtual environment that evolves over ?ve years from a human/human role-play exercise into a fully digital, AI-controlled human/computer graphical environment. An intelligent MIQA-based author agent will be evaluated on a corpus of ?ve years of recorded interactions and will achieve human levels of agency and structure (measured both subjectively and objectively) in a shared authorship Turing Test. This environment will facilitate establishing the Interactive Narrative Competition to evaluate the MIQA agent against other approaches. The event will engage students, industry, and the public in intelligent interactive narrative research, a topic that receives much interest but rarely transitions from a laboratory setting due to the high barrier to entry and lack of accessible hands-on applications.
|Effective start/end date||6/15/22 → 5/31/27|
- National Science Foundation: $119,434.00
Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.