EAGER: Planning Believable Narratives by Modeling Agent Beliefs

这个项目的目标是改进为模拟训练和教育游戏等虚拟环境自动生成故事的软件。具体来说，该软件将能够推断出什么是真实的，每个角色认为什么是真实的，他们认为别人认为什么是真实的，等等，以改善虚拟角色的行为方式，使他们看起来更可信，更人性化。目前设计这些叙事的方法通常假设代理人了解他人的信仰和目标；这通常会导致代理的行为不一致或令人难以置信，从而损害软件的可信度和人类用户的体验质量。该提案将扩展首席研究员现有的叙事计划系统，使用一种方法，让代理考虑与自己和他人的行为一致的多套信念，排除代理的信念与他们的行为不一致的情况。与现有的方法相比，这将允许叙述计划者生成更广泛的、对人类更可信的叙述，以及处理诸如欺骗和不确定的情况，其中关于信念的推理是至关重要的。研究小组将通过几个实验来测试软件和这些假设，这些实验要求人们将新软件生成的叙述与最先进的方法生成的叙述进行比较。如果成功，该项目将为提高虚拟代理与人类交互的系统的质量奠定基础，例如智能电话助手、在线游戏、自动客户聊天工具和教育软件。特别是，这项工作将导致训练场景，在这些场景中，理解他人的信仰是至关重要的，比如官员与公民的互动。这项工作也是跨学科的，从计算机科学到心理学，首席研究员致力于培养年轻研究人员从事跨越这些智力界限的工作，并招募那些可能不参与计算机科学相关研究的研究人员。在这项工作中，首席研究员提出了一种基于随机模态逻辑和可能世界推理的智能体信念模型，适用于协调虚拟环境的规划算法。通过支持单个模态“相信”谓词，规划器可以将叙事搜索空间视为Kripke结构来推理认知上可访问的状态。这改进了以前的模型，允许任意嵌套的信念，同时减少了虚拟环境作者编写替代场景的负担，从而增加了它们的灵活性和表现力。研究小组将把这种信念模型整合到一个原型系统中，该原型系统基于之前由首席研究员开发的Glaive叙事规划器。这个原型将利用Glaive现有的启发式驱动的状态空间搜索技术：除了扩展暂时可访问的状态，Glaive还将扩展认知可访问的状态，并跟踪在Kripke结构中，认知子节点所采取的行动何时可以被其认知父节点所预测。最初的原型对于实时使用来说太慢了，但它将适合进行所提出的实验，研究这种模型在多大程度上提高了自动生成故事中智能体行为的可信度。特别是，该团队将研究计划者是否产生的叙事结构更符合人类观众的期望：也就是说，模型将更类似于人类观众回答有关代理信念的问题，而最终计划者将更像人类作者撰写的故事。此外，该原型有望解决某些缺乏代理信念模型的算法无法解决的叙事规划问题。这些主张将通过让新的原型和两个最先进的计划者为团队开发的场景库生成叙事来进行评估，这些场景库依赖于对其他代理具有心智理论的代理，然后向系统和人类用户询问有关生成的叙事和代理信念的一些问题，以评估计划者的输出与人类期望和可信度的一致性。

The goal of this project is to improve the software that generates stories automatically for virtual environments like training simulations and educational games. Specifically, the software will be able to reason about what is actually true, what each character thinks is true, what they think others think is true, and so on, to improve the way virtual characters act and make them seem more believable and more human. Current approaches to designing these narratives often assume agents know everything about others' beliefs and goals; this often leads to inconsistent or un-believable behaviors by the agents, which damage the credibility of the software and quality of the experience for their human users. The proposal will extend the lead researcher's existing narrative planning system, using an approach that lets agents consider multiple sets of beliefs that are consistent with their own and others' actions so far, ruling out situations where agents have beliefs that are inconsistent with their actions. Compared to existing approaches, this should allow the narrative planner to generate a wider variety of narratives that are also more believable to humans, as well as to handle situations such as trickery and uncertainty where reasoning about beliefs is crucial. The research team will test the software and these assumptions through several experiments that ask people to compare narratives generated by the new software to those generated by state of the art methods. If successful, the project sets the stage to improve the quality of systems where virtual agents interact with humans such as smart phone assistants, online games, automated customer chat tools, and educational software. In particular, the work will lead to training scenarios where understanding others' beliefs is crucial, such as officer-citizen interactions. The work is also interdisciplinary, ranging from computer science to psychology, and the lead researcher is committed to training young researchers to do work that crosses these intellectual boundaries and to recruiting researchers who might not otherwise participate in computer science-related research.In the work, the lead researcher proposes to develop a model of agent belief based on doxastic modal logic and possible worlds reasoning suitable for use in a planning algorithm that coordinates a virtual environment. By supporting a single modal 'believes' predicate, the planner can treat the narrative search space as a Kripke structure to reason about epistemically accessible states. This improves on previous models by allowing arbitrarily nested beliefs while simultaneously reducing the burden on the virtual environment's author to write alternative scenarios, thus increasing their flexibility and expressiveness. The research team will integrate this model of beliefs into a prototype system based on the Glaive narrative planner previously developed by the lead researcher. This prototype will take advantage of Glaive's existing heuristic-driven state-space search techniques: in addition to expanding temporally accessible states, Glaive will also expand epistemically accessible states and track when an action taken by an epistemic child can be anticipated by its epistemic parent in the Kripke structure. The initial prototype will be too slow for real-time use, but it will be suitable for conducting the proposed experiments that investigate to what extent such a model improves the believability of agent behavior in automatically generated stories. In particular, the team will study whether the planner produces narratives whose structure better meets the expectations of a human audience: that is, the model will answer questions about agent beliefs more similarly to a human audience and the resulting planner will generate stories more like those composed by human authors. Further, the prototype is expected to solve certain narrative planning problems which algorithms that lack a model of agent beliefs cannot solve. These claims will be evaluated by having the new prototype and two state-of-the-art planners generate narratives for a library of scenarios to be developed by the team that rely on agents having a theory of mind for other agents, then asking both the systems and human users a number of questions about the generated narrative and agents' beliefs to evaluate how well the planners' output conforms with humans' expectations and believability.