Abstract
In automated story generation, creating believable characters requires a delicate balance between persistence and adaptability. This paper introduces a formal model for "intention revision" that bridges Belief-Desire-Intention (BDI) logic with Partial-Order Causal Link (POCL) planning. By defining specific triggers for reconsideration—namely, causal link threats—the framework ensures that characters only abandon goals when they become truly unachievable. Furthermore, the authors propose the QUEST cognitive model as a metric for evaluating whether these character revisions remain explainable and consistent to a human audience.
Introduction
A central challenge in narrative AI is the modeling of character intentions. Previous approaches often resulted in agents at two extremes: "Conscious Rationalizers," who are overly rigid and rarely deviate from a plan, and "Fairweather Agents," who drop goals at the first sign of a minor complication. Neither extreme captures the nuance of human-like behavior, where persistence is tempered by rational adaptation. To address this, we propose a formalization of intention revision that allows characters to change their minds in a way that is both logically sound and narratively satisfying.
Formalizing Intention Revision
Our methodology integrates two distinct AI paradigms: BDI agents and POCL planning. The core of the model is the Reconsidered Intention Frame. Unlike traditional agents that might re-evaluate their entire goal set at every step, our model uses specific "causal link threats" as triggers. A threat occurs when an observation in the story world breaks a prerequisite for a planned action.
Drawing on Cohen and Levesque’s theory of persistent goals (P-GOAL), the agent maintains a commitment to its intention until it believes the goal has been achieved or has become impossible. When a goal is deemed unachievable due to a causal threat, the agent performs a revision, dropping the old intention and adopting a new, achievable one that aligns with the broader narrative requirements.
Evaluating Explainability via QUEST
To ensure that these revisions are perceived as "believable" by an audience, we utilize the QUEST cognitive model. QUEST maps narrative elements into a Knowledge Structure (QKS)—a directed graph where nodes represent events, states, and goals, and edges represent causal or intentional relationships.
By analyzing the QKS, we can mathematically predict how humans will answer "Why?" and "How?" questions regarding a character's change of heart. This allows for a quantitative measurement of the Goodness of Answer (GOA). If a character's revision follows the paths defined in the QKS, the audience is more likely to find the behavior explainable.
Key Insights and Results
The primary contribution of this work is the set of formal mathematical definitions (Definitions 1-5) that provide a rigorous basis for intention revision within POCL planning. Key insights include:
- Causal Threats as Triggers: Using plan failures as catalysts for reflection avoids the computational overhead of constant re-evaluation and mirrors narrative tension.
- Persistence via P-GOAL: Characters exhibit realistic commitment, giving up only when a goal is logically impossible based on their current beliefs.
- Cognitive Validation: By aligning AI behavior with the QUEST model, we provide a roadmap for creating characters whose internal shifts feel earned rather than arbitrary.
Future Work
As a work-in-progress, several avenues for expansion remain. Future research will focus on the algorithmic "plan surgery" required to seamlessly integrate new intentions into existing story structures. Additionally, we plan to conduct human subject trials to empirically validate that the QUEST model’s predictions of explainability match human intuition in complex revision scenarios. Finally, moving beyond ground literals to more complex, uncertain belief systems will further enhance character depth.
Summary updated