NSF-BSF: RI: Small: Decentralized Active Goal Recognition

NSF-BSF：RI：小型：去中心化主动目标识别

• 批准号: 1816382
• 负责人: Christopher Amato
• 金额: $ 47万
• 依托单位: Northeastern University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2018
• 资助国家: 美国
• 起止时间: 2018-07-01 至 2023-12-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1816382&HistoricalAwards=false
• 关键词: NSF; BSF; RI; Small; Decentralized

Autonomous systems often need to coordinate with other sensors, robots, autonomous cars, and people. This results in multi-agent systems, in which agents must be able to determine what others are currently doing and predict what they will be doing in the future. This task of plan and goal recognition, typically relies upon a passive observer that continually observes the multi-agent system. In many real-world systems, such as assistive robotics in the home, this is not practical. Real-world systems will require active goal recognition, where information has a cost, and other tasks are pursued and completed continuously during goal recognition. For example, consider a team of robots assisting a disabled or an elderly person. The robots must fetch items and clean areas, while also opening doors or otherwise escorting the person. The agents will have to balance completion of their own tasks with information gathering about the target person's behavior. Current goal recognition methods cannot solve this active goal recognition problem. Furthermore, in realistic multi-agent domains including agricultural applications, disaster assistance, or military settings, communication will be limited or noisy. This will require decentralized active goal recognition methods where agents make choices based on their own limited viewpoints. Developing such active goal recognition methods will be the focus of this research. More specifically, the research will develop new methods for active goal recognition to allow teams of agents to coordinate with other systems. The project will develop methods for: active goal recognition, combining the observer's planning problem with goal recognition to balance information gathering with task completion for a single agent (observer) and single target, decentralized active goal recognition, combining multi-agent planning for the observers with goal recognition to balance information gathering with task completion and coordination for multiple observer agents and a single target agent, and decentralized active goal recognition of multiple targets, combining multi-agent planning for the observers with goal recognition to balance information gathering with task completion and coordination for multiple observer agents and target agents. The research will develop a range of methods that are based on classical, information-theoretic and decision- theoretic planning that exploit the special structure in our problem. The work will be tested on a range of common benchmarks, against current methods and in multi-robot domains to ensure realistic experiments. This research will consider active goal recognition (combining an observer's planning problem with goal recognition of a target) in single-agent and decentralized multi-agent environments. The resulting work will greatly extend the usefulness of goal recognition, making it realistic to use in scenarios when information gathering has a cost and other tasks may need to be completed by the observer(s).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.

自主系统通常需要与其他传感器，机器人，自动驾驶汽车和人员进行协调。 这导致了多代理系统，其中代理必须能够确定其他人当前正在做什么并预测他们将来将要做什么。计划和目标识别的任务通常依赖于不断观察多代理系统的被动观察者。在许多现实世界中，例如家庭中的辅助机器人技术，这是不实际的。现实世界中的系统将需要积极的目标识别，其中信息有成本，并且在目标识别期间继续执行并连续完成其他任务。例如，考虑一个机器人团队来协助残疾人或老年人。机器人必须获取物品和清洁区域，同时还必须打开门或护送该人。代理商将必须平衡自己任务的完成与有关目标人行为的信息收集。当前的目标识别方法无法解决此主动目标识别问题。此外，在现实的多机构领域中，包括农业应用，灾难援助或军事环境，交流将是有限或嘈杂的。这将需要分散的主动目标识别方法，其中代理根据自己的有限观点做出选择。开发这种积极的目标识别方法将是这项研究的重点。更具体地说，该研究将开发新的积极目标识别方法，以允许代理团队与其他系统协调。该项目将开发出：积极目标识别的方法，将观察者的计划问题与目标识别相结合，以平衡信息收集与单个代理的任务完成（观察者）和单个目标，单个目标，分散的主动目标识别，将观察者的多代理计划与目标识别与多个目标识别的多代理计划与多个目标的识别和多个观察者的识别，并与多个观察力的目标进行平衡，并组合多个观察者，并组合多个观察者，并组合多个观察者，并组合了单个观察者的单个目标代理和组合单个目标代理人和分数分数的组合。具有目标识别的观察者的多代理计划，以平衡信息收集与多个观察者和目标代理的任务完成和协调。该研究将开发一系列基于经典，信息理论和决策理论计划，这些方法利用了我们问题中的特殊结构。这项工作将在一系列常见基准测试中，以当前方法和多机器人域进行测试，以确保实验实验。这项研究将考虑单个和分散的多机构环境中的主动目标识别（将观察者的计划问题与目标识别相结合）。最终的工作将极大地扩展目标识别的有用性，这使得在信息收集的成本和其他任务可能需要由观察者完成时，在方案中使用了现实。该奖项反映了NSF的法定任务，并被认为是通过基金会的知识分子和更广泛影响的评估来评估的支持。

项目成果

Hybrid Independent Learning in Cooperative Markov Games

• DOI: 10.1007/978-3-030-64096-5_6
• 发表时间: 2020
• 作者: Roi Yehoshua;Chris Amato

Unbiased Asymmetric Reinforcement Learning under Partial Observability

• DOI: 10.5555/3535850.3535857
• 发表时间: 2021-05
• 作者: Andrea Baisero;Chris Amato

A Deeper Understanding of State-Based Critics in Multi-Agent Reinforcement Learning

• DOI: 10.1609/aaai.v36i9.21171
• 发表时间: 2022-01
• 作者: Xueguang Lyu;Andrea Baisero;Yuchen Xiao;Chris Amato

Belief-Grounded Networks for Accelerated Robot Learning under Partial Observability

• 发表时间: 2020-10
• 期刊: ArXiv
• 作者: Hai V. Nguyen;Brett Daley;Xinchao Song;Chris Amato;Robert W. Platt

Equivariant Reinforcement Learning under Partial Observability

• 发表时间: 2023
• 作者: Hai Huu Nguyen;Andrea Baisero;David Klee;Dian Wang;Robert Platt;Christopher Amato