CAREER: Structured High-Agency Interactive Narratives for Virtual Environments

该奖项的全部或部分资金来自《2021年美国救援计划法案》(公法117-2)。叙事是我们思考、交流和学习的基本方式。虚拟环境，如训练模拟，邀请用户扮演叙事中的一个角色，而系统控制所有其他非玩家角色和环境。互动叙事是教人们如何完成任务和就重要话题教育人们的有效工具，但撰写互动叙事是一项具有挑战性的工作。大多数都是手动编写的，这确保了良好的结构，但限制了它们的范围，因为每个选择都必须事先想象好。有些环境是逼真的模拟，让用户可以自由地做各种各样的动作，但设计师很难保证故事有必要的内容。该项目将使用人工智能规划算法在游戏和训练模拟中的运行时创建叙事。有计划的叙事可以有手写故事的结构和模拟的自由。这个项目将探索用于生成叙事以及用户记忆和期望的模型的快速算法。在项目持续时间内，研究团队将开发虚拟环境(如针对警察的虚拟现实降级培训模拟)，从两个人之间的角色扮演练习演变为完全自动化的虚拟环境，人工智能在其中为每个玩家个性化互动叙事。该项目将互动叙事框定为控制一个角色的玩家和控制虚拟环境所有其他元素的体验经理之间的即兴练习。这些伙伴通过他们选择采取的行动来交流他们的信念、意图、记忆和期望，这是一个需要推理才能理解的嘈杂渠道。该项目将交互式叙事创作作为一个相互隐含的问题回答(MIQA)过程来操作。一个参与者采取的每一项行动都会导致他们的伴侣含蓄地询问他们为什么要采取该行动和/或含蓄地回答先前提出的问题。一方对另一方提出的问题回答得越好，他们就越接近相互理解。MIQA结合了对多智能体人工智能规划、记忆和预期的认知模型以及自动提问回答过程的研究，以代表双方以及他们相互理解的程度。在交互式虚拟环境中，参与者将进行配对练习，其中一个将扮演玩家，另一个将担任体验经理。在这次练习中，双方都将回答有关他们的行为以及他们对伴侣行为的看法的问题。他们还将报告他们对叙事结构和球员经纪公司的看法。这些练习将从个人对个人的练习开始，但通过数据收集和模型改进将演变为个人对智能代理的练习。研究假设是，智能代理将能够提供高代理、结构化的交互叙事，其质量接近与人类合作伙伴创建的叙事质量。这一假设将使用图灵测试进行评估：玩家能否识别交互叙事是由人类还是智能代理控制的。这些演习将在一个名为Camelot的虚拟环境中进行，但研究团队将同时在正在进行的虚拟现实培训模拟中实施相同的程序，该模拟正在与警察培训专家协商建立，以教授降低潜在危险情况的最佳实践。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).Narratives are fundamental to the way we think, communicate, and learn. Virtual environments such as training simulations invite the user to play the role of one character in a narrative, while the system controls all the other non-player characters and the environment. Interactive narratives are effective tools for teaching people how to perform a task and educating people about important topics, but writing interactive narratives is challenging. Most are manually written, ensuring a nice structure but limiting their scope because every choice must be imagined in advance. Some environments are realistic simulations that give users freedom to do a wide variety of actions, but then it is hard for the designer to guarantee the narratives have the necessary content. This project will use artificial intelligence planning algorithms to create narratives at run time in games and training simulations. Planned narratives can have the structure of a hand-written story and the freedom of a simulation. This project will explore fast algorithms for generating narratives as well as models of what users remember and expect. Over the project duration, the research team will develop virtual environments (such as a virtual reality de-escalation training simulation for police officers) that evolves from a role-playing exercise between two people to a fully automated virtual environment where the artificial intelligence personalizes the interactive narrative for each player.This project frames interactive narratives as an improvisational exercise between a player who controls one character and an experience manager who controls all the other elements of a virtual environment. These partners communicate their beliefs, intentions, memories, and expectations via the actions they choose to take, which is a noisy channel that requires inference for understanding. This project operationalizes the interactive narrative creation as a Mutual Implicit Question Answering (MIQA) process. Each action taken by one participant causes their partner to implicitly ask questions about why they took that action and/or implicitly answers questions that were raised earlier. The better one partner can answer questions raised by the other, the closer they are to mutual understanding. MIQA combines research on multi-agent artificial intelligence planning, cognitive models of memory and expectations, and procedures from automated question answering to represent both partners and how well they understand one another. Participants in an interactive virtual environment will do paired exercises where one will act as player and the other as experience manager. During this exercise, both partners will answer questions about their actions and about their perceptions of their partner’s actions. They will also report on their perceptions of the structure of the narrative and the player agency. These exercises will begin as person-to-person exercises but will evolve into person-to-intelligent-agent exercises through data gathering and model refinement. The research hypothesis is that the intelligent agent will be able to provide high-agency, structured interactive narratives that approach the quality of those created with a human partner. This hypothesis will be evaluated using a Turing test: can the player identify whether the interactive narration is controlled by a human or an intelligent agent. These exercises will take place in a virtual environment called Camelot, but the research team will simultaneously implement the same procedures into an ongoing virtual reality training simulation that is being built in consultation with police officer training experts to teach best practices for de-escalating potentially dangerous situations.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.