Collaborative Research: III: Medium: VirtualLab: Integrating Deep Graph Learning and Causal Inference for Multi-Agent Dynamical Systems

协作研究：III：媒介：VirtualLab：集成多智能体动态系统的深度图学习和因果推理

• 批准号: 2312501
• 负责人: Yizhou Sun
• 金额: $ 80万
• 依托单位: University of California-Los Angeles
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-08-15 至 2026-07-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2312501&HistoricalAwards=false
• 关键词: Collaborative; Research; III; Medium; VirtualLab; Collaborative; Research; III; Medium; VirtualLab

Many real-world domains, spanning physical systems, social systems, brain networks, and epidemic networks, can be conceptualized as multi-agent dynamic systems, wherein different agents interact with each other and progress according to specific dynamics. Understanding and modeling these systems can enhance our comprehension of their underlying mechanisms, allowing us to make more accurate long-term predictions and better-informed decisions, with or without interventions. Despite extensive study of multi-agent dynamical systems in specific domains, there is currently no general solution available, and even the most knowledgeable experts may struggle to describe them mathematically. The proposed VIRTUALLAB framework aims to create a virtual lab capable of learning system dynamics from observed data, predicting future agent trajectories, and accurately forecasting potential system outcomes under a range of interventions. This project will facilitate the rapid adoption of AI techniques in different domains, promoting the digital revolution and the use of AI for healthcare, science, and public policy. The investigators plan to incorporate educational activities into the research, offering students exciting opportunities to apply AI and ML in various domains such as biomedical research, material science, and public health. They will also widely disseminate their findings through publications, tutorials at various conferences, and collaborations with domain experts. The project has identified several limitations in existing approaches to modeling and predicting multi-agent dynamical systems. Firstly, approaches are often domain specific, and there is a lack of general methodology to address the full range of dynamical systems. Secondly, most dynamical systems are defined by complex ordinary or partial differential equations that can be difficult or even impossible to devise. Thirdly, making predictions can be very time-consuming and may not be applicable to large-scale systems. Lastly, very little work has addressed the problem of causal inference in multi-agent systems. The VIRTUALLAB framework is designed to be transformative and address these challenges. Firstly, it will provide general solutions to model multi-agent dynamical systems across a broad spectrum of applications, where the dynamics can be learned from incomplete and irregular observational data from the same or related systems. This will involve addressing several challenges, such as modeling continuous dynamics from incomplete signals, designing models that capture high-order nonlinear dynamics, generalizing learned dynamics to new systems with few observations, and scaling models to handle large-scale systems and make training and inference efficient for real-world systems. Secondly, VIRTUALLAB will provide accurate predictions of potential outcomes after an intervention, either at the node or system level, by leveraging offline data. Doing so will involve handling both system- and node-level intervention and continuous-time dynamic intervention, rather than static intervention that may occur in the future. Lastly, the project will test and evaluate the proposed framework using several use cases, including functional brain networks, molecular dynamics, and epidemic dynamics.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.

许多现实世界的域，跨越物理系统，社交系统，大脑网络和流行网络，可以概念化为多代理动态系统，其中不同的代理相互交互并根据特定的动态进行进展。理解和建模这些系统可以增强我们对它们的基本机制的理解，从而使我们能够在有或没有干预的情况下做出更准确的长期预测和更有信息的决策。尽管对特定领域中的多代理动力系统进行了广泛的研究，但目前尚无通用解决方案，即使是最知识渊博的专家也可能难以在数学上描述它们。提出的Virtuallab框架旨在创建一个虚拟实验室，能够从观察到的数据中学习系统动态，预测未来的代理轨迹，并准确预测在一系列干预措施下的潜在系统结果。该项目将促进在不同领域中快速采用AI技术，从而促进数字革命并将AI用于医疗保健，科学和公共政策。调查人员计划将教育活动纳入研究中，为学生提供激动人心的机会，将AI和ML应用于生物医学研究，材料科学和公共卫生等各个领域。他们还将通过出版物，各种会议的教程以及与域专家的合作来广泛传播他们的发现。该项目已经确定了建模和预测多代理动力系统的现有方法中的几个局限性。首先，方法通常是特定领域的，并且缺乏解决全部动态系统的一般方法。其次，大多数动态系统是由复杂的普通或部分微分方程定义的，这些方程可能难以设计甚至不可能设计。第三，做出预测可能非常耗时，并且可能不适用于大型系统。最后，很少的工作解决了多代理系统中因果推断的问题。 Virtuallab框架旨在具有变革性并应对这些挑战。首先，它将提供一般解决方案，以模拟各种应用程序的多代理动力系统，在这些应用程序中，可以从相同或相关系统的不完整和不规则的观察数据中学到动态。这将涉及解决几个挑战，例如从不完整的信号中建模连续动态，设计捕获高阶非线性动态的模型，将学习的动态推广到具有很少的观测值的新系统，以及缩放模型来处理大规模系统，并使培训和推理对现实世界系统进行培训和推理效率。其次，通过利用离线数据，Virtuallab将在节点或系统级别进行干预后的潜在结果进行准确的预测。这样做将涉及处理系统和节点级干预以及连续的时间动态干预，而不是将来可能发生的静态干预。最后，该项目将使用多种用例来测试和评估所提出的框架，包括功能性大脑网络，分子动力学和流行动力学。该奖项反映了NSF的法定任务，并被认为是值得通过基金会的知识分子优点和更广泛影响的审查标准来评估的支持。