Emergent Narrative Generation with Social Planning Agents

以社會規劃代理人進行茁現式敘事生成

張學民(Hsueh-Min Chang)

指導教授 ： 蘇豐文

國立清華大學/電機資訊學院/資訊工程學系/博士(2010年)

摘要

Computer-based interactive storytelling is an inherently contentious subject both as an emerging genre of art /entertainment and as a research topic. On one hand, the computer media endows agency to the audience, transforming their members from passive viewers to active participants. On the other hand, the freedom of user to do anything jeopardizes story quality by disrupting authorial control over the overall plot. This dilemma leads to a variety of AI-based interactive storytelling systems. Among the variety, emergent narrative is an agency-centric approach that seeks to generate stories from autonomous interaction among agent-controlled non-player characters (NPCs) and/or the human player. However, narrative research has not yet shown conclusive evidence that a set of fully autonomous agents can produce event sequences that qualify as a story at all. Two of the most important problems are the lack of a mental dimension and narrative non-emergence. First of all, characters’ consideration on the minds of other characters is a key element underpinning what is known as a story. NPC agents that simply pursue their goals without a model of others' beliefs and intentions cannot reflect this mental dimension. Second, even if the NPCs think and act socially, the initial state of the game setting may not yield a sufficiently long story or anything at all because characters may be unable to find any possible activity, and they may conclude their activities too quickly even if they find some.
My dissertation attacks both problems by designing agents who solve problems by influencing the minds and actions of other agents. Such a mode of problem solving is designated as social planning. To capture the psychological factors in narratives, social planning agents are designed to possess theories of mind, which are their models of the psychology of other characters. A social planning agent then reasons about the minds of other agents, how they update beliefs and how they generate goals. Based on this social reasoning, the social planning agent plans to motivate other agents to take part in its overall plan. The story is the aggregation of the consequences of social plans. A novel planner, namely HSSP, is developed to do social planning by combining plan search and theorem proving.
I qualitatively show that HSSP produces interesting narrative segments at a mental level by presenting a sample of generated narratives which are not back-engineered from a known story. Quantitatively, I experimentally demonstrate a dilemma of narrative emergence. A narrative universe should be eventful, which means that social plans frequently happen. It should also be intriguing, which requires longer social plans. However, while a high sensitivity of NPCs' minds to external influences increases the frequency at which social plans appear, it also decreases the average social plan length. This dilemma is solved with a story facilitator named DIOS, which adjusts character personalities on the fly to make stories happen without harming the social plan length.

關鍵字

數位敘事 ； 故事產生 ； 智慧型代理程式 ； 電腦遊戲

並列摘要

茁現式(emergent)敘事產生是使用自主性的軟體代理人扮演故事角色以產生敘事的技巧。相較傳統的故事產生方法，茁現式敘事具有可和互動媒體整合的優點，並且允許使用者參與其中，改變敘事情節。然而以代理人為基礎的虛擬場景常被質疑無法有效產生敘事，主要原因有二。首先，代理人角色的行為往往缺乏心理互動層面，如此產生的內容不足以稱為敘事。其次，可能出現代理人角色無事可做，導致沒有任何情節產生的狀況。
本論文提出一個以知識為基礎的虛擬場景架構，其中核心的元件為一社會互動規劃器(social planner)。此規劃器結合自動定理證明與Heuristic Search規劃演算法，可使角色能夠對其他角色的信念與意圖做推理，進而影響其他角色的意圖與行為。在實例探討中，給予角色個性，知識及外在環境初始設定後，此系統可以產生精簡版的莎士比亞戲劇《奧賽羅》以及200多個變體。
本文以實驗證實，在茁現式系統中，敘事產生的機率與所產生的敘事長度有矛盾關係。當角色心理容易受到外在影響而變動時，平均敘事產生的機率提升，然而敘事的平均長度縮短。為有效兼顧機率與長度，本文發展出稱為DIOS的故事協助模組(story facilitator)。DIOS能以對角色初始個性影響最少的方式，線上動態修改人物個性，在不縮短敘事長度的狀況下增加敘事產生機率。

並列關鍵字

digital narrative ； story generation ； intelligent agent ； compute game

參考文獻

[1] Chang, H.-M. and Soo, V.-S. 2009. Planning-based Narrative Generation in Simulated Game Universes. IEEE Transactions on Computational Intelligence and AI in Games 1(3): 200-213.

[2] Chang, H.-M. and Soo, V.-S. 2008a. Simulation-based story generation with a theory of mind. In Proceedings of the Fourth Conference on Artificial Intelligence and Interactive Digital Entertainment (Stanford, California, U.S.A, October 22-24, 2008), 16-21, AAAI Digital Library.

[5] Chiu, C.-C., Kao, E. C.-C. Chang, P. H.-M. and Soo, V. W. 2007. AI-RPG toolkit: towards a deep model implementation for improvisational virtual drama. In Proceedings of the 7th International Conference on Intelligent Virtual Agents (Paris, France, September 17-19, 2007) 368. Lecture Notes in Computer Science 4722.

[7] Chang, P. H.-M., Chen, K.-T., Chien, Y.-H., Kao, E., Soo, V.-W. 2005a. From reality to mind: a cognitive middle layer of environment concepts for believable agents. In Weyns, D., Parunak, H. V. D., Michel, F. (eds): Environments for Multi-Agent Systems, 57-73, Lecture Notes in Computer Science 3374.

[Referenced Publications]