Simulation-Based Story Generation with a Theory of Mind

Emergent narrative refers to simulation-based systems in which stories emerge from the autonomous interactions among character agents and/or the human player. Despite its advantages in interactive games, emergent narrative attracts concern about whether complex narratives emerge from arbitrary interactions. This problem can be alleviated if the character agents act according to what a story should require. Based on the observation that deliberate influence on others' minds constitutes the fabric of most stories, we propose that the capability for social influence is crucial for narrative agents. This paper presents a specialized planning technique called social planning, which allows a character agent to achieve its goal by reasoning about other characters' minds and influencing their actions. A prototype system based on social planning agents succeeds in generating simplified variations of Shakespeare's Othello through simulation. Introduction: Games and Story Generation Games are a space in which the player instantiates stories (Frasca 1999) with varying degree of freedom. Player agency is said to be a central feature of the computer as a new media form (Murray 1997). Depending on the nature of the game and ways of playing, each game session can be a unique narrative. A game has high user agency if it gives the player the feeling of being able to control the progression. Assuming that user agency is desirable, storytelling systems should attempt to increase the range of possible variations across game sessions. This has been done in AI systems by employing characters that uses plans (Pizzi and Cavazza 2007), emotional responses (Aylett et al. 2006) and decision-theoretic reasoning (Si, Marsella, and Pynadath 2005) to autonomously respond to various inputs. Nevertheless, storytelling systems typically seek to maintain a consistent backbone story from which the character agents should never deviate too far. In practice, if the deviate too far they risk invalidating the backbone story and making proceeding impossible. The emergent approach toward story generation has the advantage of being naturally compatible with user agency. In an ideal emergent narrative system, autonomous interactions between characters should constantly make up Copyright © 2008, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. story-quality content as the player goes along, as in the words of (Aylett et al. 2006). The main concern, however, is that the quality of story content resulting from arbitrary interactions is at least questionable. On the other hand, story generators seldom assume a backbone story. They simply take a collection of data, often consisting of an initial state of the world, and output stories that contain action sequences and other events. Each initial state thus could lead to a unique story, yet not all possible action sequences are equally story-like. Story generators tend to have a set of criteria about what action sequences should be regarded as stories. Among them, social influence between characters is considered a central feature. For example, TALESPIN (Meehan 1977), an early story generator, puts much emphasis on beliefor intention-influencing actions. The IPOCL story planner (Riedl and Young 2004) also emphasizes social influence by ensuring that each action is backed with an intention, which is usually the effect of another character's action. If autonomous characters in an emergent narrative are able to act in accordance with some criteria about what a story should be like, the result can be comparatively story-like. In this paper, we take the issue of social influence as central, and propose giving character agents the reasoning machinery to exert social influence, or to achieve personal goals through social influence. The character agents plan not only to manipulate physical objects, but also to manipulate the minds of other agents. Simulation based on such agents can produce action sequences that resemble those that are generated by a story planner. This paper reports our experience in the design and implementation of such a story simulation system, as a step towards a both open-ended and storied game experience. Social Influence in Stories From a biological standpoint, Dautenhahn (2003) views stories as the retelling of events in a social group, where emotions and intentions of group members are reconstructed. Social influence plays a major part in Bremond's narratology work, which is based on distinction between agents 1 and patients; the former role exerts 1 We are forced to use three related but different senses of Proceedings of the Fourth Artificial Intelligence and Interactive Digital Entertainment Conference