RI: Small: Sound Abstractions for Efficient and Reliable Automated Planning

RI：小型：高效可靠的自动化规划的健全抽象

• 批准号: 1909370
• 负责人: Siddharth Srivastava
• 金额: $ 49.99万
• 依托单位: Arizona State University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-08-01 至 2023-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1909370&HistoricalAwards=false
• 关键词: RI; Small; Sound; Abstractions; Efficient

Autonomous systems have the potential to transform society by assisting people in activities ranging from household elder care to space exploration. Such agents need to be versatile enough to solve a range of tasks without prior task-specific programming. However, it is difficult to compute the behavior that an autonomous agent should execute in order to achieve a new objective (e.g., to search for rocks with high moisture content). This project will develop a new framework and algorithms that utilize the principle of abstraction to efficiently compute task-specific behavior for autonomous agents in a domain-independent fashion. Since abstraction blurs detail in general, this principle has been difficult to employ in practice -- plans computed using abstract models can miss details, and can be dangerous to use with real-world autonomous systems. This project will develop new methods that keep track of imprecision created by abstraction and effectively resolve it, to efficiently compute reliable plans.More precisely, this project will develop a formal framework for analyzing and creating abstract models that are sound, i.e., they permit only correct inferences (and possibly a subset of correct inferences) with regard to the most accurate model available. In general, such models distinguish model imprecision caused by abstraction from non-determinism or stochasticity inherent in the domain. It will develop new paradigms for efficient planning using sound abstract models while remaining provably correct with regard to the most accurate model available. These planning paradigms address settings with and without uncertainty; they will also allow AI systems to compute complex plans by automatically composing planners that are independently designed to be efficient for different types of abstract models.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.

自主系统通过帮助人们从事从家庭养老到太空探索等各种活动，有可能改变社会。这样的代理需要足够通用，以解决一系列任务，而无需事先进行特定于任务的编程。然而，为了实现一个新的目标（例如，寻找高含水量的岩石），很难计算一个自主代理应该执行的行为。该项目将开发一个新的框架和算法，利用抽象原理以独立于领域的方式有效地计算自主代理的特定任务行为。由于抽象通常会模糊细节，因此这一原则很难在实践中应用——使用抽象模型计算的计划可能会遗漏细节，并且在现实世界的自治系统中使用可能是危险的。该项目将开发新的方法，跟踪抽象产生的不精确并有效地解决它，以有效地计算可靠的计划。更准确地说，这个项目将开发一个正式的框架，用于分析和创建可靠的抽象模型，也就是说，它们只允许对最准确的可用模型进行正确的推断（可能是正确推断的子集）。一般来说，这些模型将抽象引起的模型不精确与领域固有的非确定性或随机性区分开来。它将开发新的范例，使用合理的抽象模型进行有效的规划，同时保持可证明的关于最准确的可用模型的正确性。这些规划范例处理有或没有不确定性的环境；它们还将允许人工智能系统通过自动组合计划器来计算复杂的计划，这些计划器是独立设计的，可以有效地处理不同类型的抽象模型。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

Differential Assessment of Black-Box AI Agents

黑盒人工智能代理的差异化评估

• DOI: 10.1609/aaai.v36i9.21223
• 发表时间: 2022
• 期刊: Proceedings of the AAAI Conference on Artificial Intelligence
• 作者: Nayyar, Rashmeet Kaur;Verma, Pulkit;Srivastava, Siddharth
• 通讯作者: Srivastava, Siddharth

Using Deep Learning to Bootstrap Abstractions for Hierarchical Robot Planning

• DOI: 10.5555/3535850.3535982
• 发表时间: 2022-02
• 作者: Naman Shah;Siddharth Srivastava

Using Deep Learning to Bootstrap Abstractions for Robot Planning

使用深度学习引导机器人规划的抽象

• 发表时间: 2022
• 期刊: Proceedings of the International Conference on Autonomous Agents
• 作者: Shah, Naman;Srivastava, Siddharth
• 通讯作者: Srivastava, Siddharth

Learning Generalized Relational Heuristic Networks for Model-Agnostic Planning

学习广义关系启发式网络以进行与模型无关的规划

• 发表时间: 2021
• 期刊: Proceedings of the AAAI Conference on Artificial Intelligence
• 作者: Rushang Karia, Siddharth Srivastava

Learning Generalized Relational Heuristic Networks for Model-Agnostic Planning.

学习广义关系启发式网络以进行与模型无关的规划。

• 发表时间: 2021
• 期刊: Proceedings of the AAAI Conference on Artificial Intelligence
• 作者: Rushang Karia, Siddharth Srivastava.