CAREER: Toward Artificial General Intelligence for Complex Adaptive Systems: A Natural Concurrent “Learning-in-Learning” Control Paradigm

职业：走向复杂自适应系统的通用人工智能：自然并发“学习中学习”控制范式

• 批准号: 2047064
• 负责人: Zhen Ni
• 金额: $ 50.01万
• 依托单位: Florida Atlantic University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-03-15 至 2026-02-28
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2047064&HistoricalAwards=false
• 关键词: CAREER; Toward; Artificial; General; Intelligence

Artificial intelligence (AI) technologies are transforming nearly every aspect of our lives and reinforcement learning (RL) is viewed as one of next big research topics in the current AI wave. While the existing AI and RL achievements are exciting, the fundamental research of data aggregation, learning and approximation capability, and the performance generalization during uncertainties, is not fully yet developed. There is still a gap from the current state-of-the-art techniques to the artificial general intelligence that can bring good performance in learning speed, data efficiency, and generalization of the optimization performance.Inspired by this observation, the PI proposes a natural concurrent RL framework that carries three major advantages over traditional RL methods, namely the i) advantages of simultaneously learning multimodal properties of the complex system; ii) structural advantages of using a personalized learning scheme; and iii) implementation advantages of the data-driven sample-efficient design. Within this framework, the PI proposes to design two concurrent RL methods to consolidate past experiences and anticipatory knowledge and build the “learning-in-learning” control paradigm. The theoretical results will certify that the proposed RL framework can be deployed with high confidence for complex adaptive systems under uncertain environments. The applications on smart energy community will support the novel learning framework and theoretical results.Beyond the scientific impacts, the proposed research has broader impacts for a wide range of research disciplines including transportation, rehabilitation, and robotics. The integration of research and education activities will also positively impact the institutions regionally and nationally. A proposed workshop will bring world renown experts to engage (state college) students and young researchers with limited financial supports to attend professional conferences. The collaboration with the industry and the national laboratory provides the students the opportunity to get external training, which can lead to competitive job offers. The proposed take-home AI/RL projects will promote interactive distance learning for schools with limited research capacity (e.g., rural community college) and for students with the preference of remote studying during the current pandemic. These activities will vigorously contribute to the nation’s AI workforce development.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.

人工智能（AI）技术正在改变我们生活的几乎每个方面，强化学习（RL）被视为当前AI浪潮中的下一个重大研究课题之一。虽然现有的AI和RL成就令人兴奋，但数据聚合，学习和近似能力以及不确定性下的性能泛化的基础研究尚未完全发展。从目前最先进的技术到人工通用智能，在学习速度，数据效率和优化性能的泛化方面仍然存在差距。受此启发，PI提出了一种自然并发RL框架，该框架与传统RL方法相比具有三大优势，即i）同时学习复杂系统的多模态特性的优点; ii）使用个性化学习方案的结构优点;以及iii）数据驱动的样本高效设计的实现优点。在此框架内，PI提出设计两种并行RL方法，以巩固过去的经验和预期的知识，并建立“学习中学习”的控制范式。理论结果将证明，所提出的RL框架可以部署在不确定环境下的复杂自适应系统的高置信度。智能能源社区的应用将支持新的学习框架和理论成果。除了科学影响，拟议的研究对广泛的研究学科，包括交通，康复和机器人技术具有更广泛的影响。研究和教育活动的一体化也将对区域和国家的机构产生积极影响。一个拟议的讲习班将使世界知名的专家参与（州立大学）学生和年轻的研究人员与有限的财政支持，参加专业会议。与行业和国家实验室的合作为学生提供了获得外部培训的机会，这可能会导致有竞争力的工作机会。拟议的带回家的AI/RL项目将促进研究能力有限的学校的互动式远程学习（例如，农村社区学院）和在当前大流行期间倾向于远程学习的学生。该奖项反映了NSF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

A Modified Maximum Entropy Inverse Reinforcement Learning Approach for Microgrid Energy Scheduling

一种改进的微电网能量调度最大熵逆强化学习方法

• DOI: 10.1109/pesgm52003.2023.10252933
• 发表时间: 2023
• 期刊: IEEE Power Energy Society General Meeting
• 作者: Lin, Yanbin;Das, Avijit;Ni, Zhen
• 通讯作者: Ni, Zhen

A new deep Q-learning method with dynamic epsilon adjustment and path planner assisted techniques for Turtlebot mobile robot

• DOI: 10.1117/12.2663695
• 发表时间: 2023-06
• 作者: W. Cheng;Zhengbin Ni;Xiangnan Zhong

Approximate dynamic programming with policy-based exploration for microgrid dispatch under uncertainties

• DOI: 10.1016/j.ijepes.2022.108359
• 发表时间: 2022-05-31
• 期刊: INTERNATIONAL JOURNAL OF ELECTRICAL POWER & ENERGY SYSTEMS
• 影响因子: 5.2
• 作者: Das, Avijit;Wu, Di;Ni, Zhen
• 通讯作者: Ni, Zhen

A Neural-Reinforcement-Learning-based Guaranteed Cost Control for Perturbed Tracking Systems

基于神经强化学习的扰动跟踪系统保证成本控制

• DOI: 10.1109/tai.2023.3346334
• 发表时间: 2023
• 期刊: IEEE Transactions on Artificial Intelligence
• 作者: Zhong, Xiangnan;Ni, Zhen
• 通讯作者: Ni, Zhen

Multi-Virtual-Agent Reinforcement Learning for a Stochastic Predator-Prey Grid Environment

• DOI: 10.1109/ijcnn55064.2022.9891898
• 发表时间: 2022-07
• 期刊: 2022 International Joint Conference on Neural Networks (IJCNN)
• 作者: Yanbin Lin;Z. Ni;Xiangnan Zhong