Advanced Techniques for Action Model Solicitation, Verification, and Induction

自主系统正在成为日常生活的重要组成部分。能够对世界进行推理并采取目标导向行动的系统处于复杂自主行为的前沿，构建它们仍然是一个挑战。人工智能的子领域被称为自动规划（AP），重点是如何构建与环境（虚拟和物理）交互的自主系统。在给定环境模型的情况下，AP通过综合代理要遵循的计划或策略来实现这一点。近几十年来，AP研究取得了巨大的进步，导致了许多行业应用，包括机器人技术、对话生成和业务流程自动化。本提案旨在直接解决部署AP解决方案的一些最大障碍。使用AP技术的一个关键瓶颈是模型获取任务：获取描述自治系统行为的规范或模型的过程。拟议的研究直接集中在这一关键的和研究不足的步骤。特别是，三个互补的研究线索将解决模型获取的各个方面——从手动模型征集到完全自主的模型归纳。第一个线程将通过将现代人工智能推理技术应用于流程来提高手动模型规范的能力。一套高级的分析和验证技术，例如自动化的基于案例的操作可达性或冗余性测试，将识别并突出模型的不一致和不足。研究的第二条主线将通过引入通用的基于逻辑的规范来统一模型表示。最后一个线程将关注模型归纳：以数据驱动的方式对AP模型进行半自主或完全自主的推理。将现代机器学习技术与AP理论的先天先验相结合，可以大大减轻从业者合成自主系统的负担。这三个线程之间的协同将导致模型获取过程中的实质性改进。工作的所有方面都反映了将形成模型获取的单一框架的综合方法。这将公开提供给专注于AP技术的研究团体，以及在当今工业中部署AP解决方案的从业者。改进的模型获取将对对话代理设计和业务流程自动化的业务领域产生直接和实质性的影响，这两者都依赖于操作模型获取。为这项研究培养高素质的人才将为他们提供在人工智能的许多基础和新兴领域发展技能的机会。我预计将有2名博士生、2名硕士生和4名本科生通过这个项目获得培训；让他们为学术界或工业界的人工智能高需求领域做好准备。

Autonomous systems are becoming a vital part of everyday life. Systems that can reason about the world and take goal-directed actions are at the forefront of sophisticated autonomous behaviour, and constructing them remains a challenge. The sub-field of Artificial Intelligence known as Automated Planning (AP) focuses on how to build autonomous systems that interact with environments (both virtual and physical). AP does so by synthesizing plans or policies for an agent to follow, given a model of the environment. AP research has made tremendous strides in recent decades, leading to numerous industry applications including robotics, dialogue generation, and business process automation. This proposal aims to directly address some of the greatest barriers to deploying AP solutions. A key bottleneck for using AP technology is the task of model acquisition: the process of acquiring specifications or models to describe the autonomous system behaviour. The proposed research directly focuses on this crucial and understudied step. In particular, three complementary threads of research will address various aspects of model acquisition -- from manual model solicitation to fully autonomous model induction. The first thread will improve the capability of manual model specification by applying modern AI reasoning techniques to the process. A suite of advanced analysis and verification techniques, such as automated case-based testing of action reachability or redundancy, will identify and highlight model inconsistencies and insufficiencies. The second thread of research will unify model representations by introducing a common logic-based specification. The final thread will focus on model induction: semi- or fully-autonomous inference of AP models in a data-driven fashion. Using a combination of modern machine learning techniques with innate priors informed by AP theory, this line of work can greatly reduce the burden on practitioners synthesizing autonomous systems. Synergy between these three threads will lead to a substantial improvement in the process of model acquisition. All aspects of the work reflect an integrated approach that will form a single framework for model acquisition. This will be made publicly available to both the research community focused on AP techniques, as well as practitioners that deploy AP solutions in industry today. Improved model acquisition will have a direct and substantial impact on the business areas of dialogue agent design and business process automation, both of which depend on action model acquisition. The training of highly qualified personnel for this research will provide them with the opportunity to develop skills in many fundamental and emerging areas of Artificial Intelligence. I expect that two PhD students, two Master's students, and four undergraduate students will acquire training through this program; preparing them for the high-demand area of AI in either academia or industry.