EAPSI: Identifying Relations between Computer-Generated and Manually Annotated Interpretations of Activities for Planning and Plan Recognition Tasks

虽然人类使用单词感知活动并通过某些属性对类似活动进行聚类，但使用无监督学习算法的计算机不一定以相同的方式识别它们，从而产生不同的解释。这项研究的目标是在人类和机器对活动的定义之间建立一种类比，这样人工智能规划和计划识别方法就不需要针对每一组定义进行调整。通常情况下，要么是人类定义的活动太模糊，机器无法进行准确的计算，要么是计算机定义的活动，使得人们无法理解潜在的推理。为每个实体开发一个解释器不仅可以在一起解决问题时平滑用户和设备之间的交互，还有助于弥合人机之间的差距。东京大学的Alex Fukunaga博士研究人工智能的自主规划、搜索、优化等多个方面，本课题将利用他感兴趣的研究领域中的技术，对以往的基于主题模型的无监督活动识别的研究进行扩展，提出了约束优化和启发式搜索两步法。第一阶段使用约束优化来将计算机识别的活动序列与给定的人类对相同序列的注释进行比对。然后，第二阶段类似地将计算机识别的活动序列与通过人类定义的分层任务网络上的启发式搜索得出的序列的注释进行比对。为了开发这些方法，该项目将包括形式化的约束和搜索问题，编码他们的配方，并测试结果的两个定义集之间的映射。本测试将通过将PI先前针对机器解释动作开发的无监督活动识别方法与使用人类感知动作表示的常用计划识别方法相结合来执行。这个NSF EAPSI奖是与日本科学促进协会合作资助的。

While humans perceive activities using words and cluster similar activities by some properties, computers using unsupervised learning algorithms do not necessarily identify them the same way and thus generate different interpretations. The goal of this research is to develop an analogy between human and machine definitions of activities so that artificial intelligence planning and plan recognition methods do not need to be adjusted for each set of definitions. Usually, either humans define the activities in a way which is too vague for machines to make accurate computations or computers define activities such that people cannot understand the underlying reasoning. Developing an interpreter for each entity will not only smooth the interaction between users and devices when solving problems together, but also contribute to bridging the human-computer gap. This project will use techniques in the research areas of interest to Dr. Alex Fukunaga at the University of Tokyo who studies many facets of artificial intelligence, especially those regarding autonomous planning, search, and optimization.Extending prior research on unsupervised activity recognition using topic models, this work proposes a two-step approach consisting of constraint optimization and heuristic search. The first phase uses constraint optimization to align a computer's recognized activity sequence with a given human's annotation of the same sequence. Then the second phase similarly aligns a computer's recognized activity sequence with an annotation of the sequence derived by heuristic search over a human-defined hierarchical task network. To develop these methods, the project will include formalizing the constraint and search problems, coding their formulations, and testing results of the identified mappings between the two definition sets. This test will be performed by combining an unsupervised activity recognition method previously developed by the PI for machine-interpreted actions with a commonly used plan recognition method that uses human-perceived action representations. This NSF EAPSI award is funded in collaboration with the Japan Society for the Promotion of Science.