CHS: Small: Strong-Story Narrative Planning for Authoring Proactive Intelligent Virtual Environments

交互式虚拟世界具有广泛的应用，包括军事训练模拟、课堂辅导系统、事件的治疗性娱乐和娱乐。这些世界邀请用户扮演一个角色，而世界及其虚拟角色的演员由系统控制。用户和系统一起创建交互式故事叙述。为了有意义并保持用户参与，用户必须自由地采取行动并看到他们选择的结果，但是手工设计交互式叙事的所有分支对于人类设计师来说很快就变得太多了。人工智能可以通过在用户做出选择时自动创建叙述来缓解这个问题。该项目将解决目前阻碍科学家创建逼真的自适应虚拟环境的几个技术限制。一个关键的限制是，探索可能的叙事空间并选择最佳叙事的算法目前太慢而不实用，特别是在人类做出不可预测的选择时。另一个原因是，设计虚拟环境的人很难确保叙事具有正确的结构并教授正确的信息。该项目将在虚拟现实培训模拟的背景下开发和测试解决这些限制的技术，帮助警察学习如何以及何时在与他们所服务的社区打交道时使用武力。虚拟现实培训提供了一种安全、经济、可重复的方式，为危险情况提供逼真和令人难忘的培训。为该项目开发的人工智能技术将确保训练模拟是真实的，无论用户做出什么选择，都能提供有效的教学。大多数以前使用人工智能控制虚拟环境的方法都集中在创建逼真的个人虚拟人。一个充满现实人物的世界可以是现实的，但设计师没有办法在叙事中强加教育或美学结构，限制了它们对训练的有用性。该项目将扩展以前对叙事规划算法的研究，以产生一个集中的叙事规划器，该规划器可以对用户，世界及其所有虚拟角色进行深入的推理，以实现与手工创作体验相同的叙事结构和质量水平，以及前面提到的非结构化环境的真实角色行为，以及开放世界环境的用户代理，玩家可以在任何时间采取任何行动。叙事规划者预测了数百万种可能的未来，并受到人类如何根据自己的信念和意图行事的模型的约束。提前规划叙事在计算上是昂贵的，但它可以在人类水平上完成，或者更好地通过采用非马尔可夫启发式搜索来解释叙事结构，只探索有希望的部分故事。然后，这些相同的算法和模型可以用来帮助虚拟环境的设计者了解他们创建的虚拟世界中什么是可能的，什么是不可能的，什么是可能的，什么是不可能的。用于模拟现实虚拟人的信念和意图模型可以用于预测和理解人类用户，当用户意外行动时，规划者预测数百万种可能性的能力可以自动调整叙事以进行补偿。这些算法将通过大量的计算实验进行验证，并最终通过测量其在真实的警察使用的自适应虚拟现实警察训练模拟中的有效性来进行验证。该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估来支持。

Interactive virtual worlds have a wide variety of applications, including military training simulations, classroom tutoring systems, therapeutic recreation of events, and entertainment. These worlds invite the user to take on the role of one character while the world and its cast of virtual characters are controlled by the system. Together, the user and the system create an interactive story narrative. To be meaningful and keep the user engaged, the user must be free to act and see the results of their choices, but designing all the branches of an interactive narrative by hand quickly becomes too much work for a human designer. Artificial intelligence can mitigate this problem by creating the narrative automatically as the user makes choices. This project will address several of the technical limitations that are currently preventing scientists from creating realistic, adaptive virtual environments. One key limitation is that algorithms for exploring the space of possible narratives and choosing the best one are currently too slow to be practical, especially in the face of the unpredictable choices that human beings make. Another is that it's hard for the people who design virtual environments to make sure the narratives have the right structure and teach the right information. This project will develop and test technologies that address these limitations in the context of a virtual reality training simulation that helps police officers learn how and when to use force when dealing with the communities they serve. Virtual reality training provides a safe, affordable, repeatable way to provide realistic and memorable training for dangerous situations. The artificial intelligence techniques developed for this project will make sure the training simulation is realistic and provides effective teaching no matter what choices the user makes.Most previous approaches for using AI to control virtual environments have focused on creating realistic individual virtual humans. A world full of realistic characters can be realistic, but there is no way for the designer to impose pedagogic or aesthetic structure on the narrative, limiting their usefulness for training. This project will extend previous research on narrative planning algorithms to produce a centralized narrative planner that reasons far into the future about the user, the world, and all of its virtual characters to achieve the same level of narrative structure and quality created by hand-authored experiences, the realistic character behavior of the previously mentioned unstructured environments, and the user agency of open world environments where the player can take any action at any time. Narrative planners anticipate millions of possible futures and are constrained by models of how humans behave according to their beliefs and intentions. Planning a narrative far in advance is computationally expensive, but it can be done at a human level or better by employing non-Markovian heuristic search that accounts for narrative structure to explore only promising partial stories. These same algorithms and models can then be used to help the designers of virtual environments understand what is possible, impossible, likely, and unlikely in the virtual worlds they create. The same models of belief and intention used to model realistic virtual humans can be used to anticipate and understand the human user, and when the user acts unexpectedly, the planner's ability to anticipate millions of possibilities can adapt the narrative automatically to compensate. These algorithms will be validated through a number of computational experiments and finally by measuring their effectiveness in an adaptive virtual reality police training simulation used by real police officers.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.