CRII: III: Informative Bayesian Learning and Data Gathering Through Expert-Acquired Data

许多实际问题中的数据都是根据用户或专家为实现特定目标而做出的决策或行动来获取的。例如，基因组学和宏基因组学数据反映了生物学家在干预过程中的策略，来自网络物理系统的数据受到专家/工程师为控制/稳定系统而做出的行动/决策的影响。许多实际系统或现象的动态性、复杂性、规模和不确定性需要最大限度地提取数据所携带的信息，以实现准确的学习/建模过程。这个项目的创新之处在于，它能够在学习过程中最优地整合用户信息，以及从非相似用户/专家获得的多个数据中学习。该项目将推进学习和数据收集过程的最新技术，并为机器学习、控制/学习理论和贝叶斯统计的科学基础做出贡献。期望在以下方面有原创性贡献：1)通过专家数据进行信息贝叶斯学习，允许从专家数据中最大限度地提取信息，用于动态或政策学习；2)多保真贝叶斯优化框架，用于可能存在大量不确定性的超大系统的有效学习；3)多专家贝叶斯学习，在学习过程中有效整合多专家数据；4)近最优贝叶斯深度强化学习数据采集，获取信息量最大的数据。对实际系统的适用性将是该项目中提出的框架的关键特征，因为它们具有严格的面向统计的性质，可以实现基于风险的、高效的、实时的和可扩展的学习和决策。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。

Data in many practical problems are acquired according to the decisions or actions made by users or experts to achieve specific goals. For instance, genomics and metagenomics data reflect the policy in the mind of biologists during the intervention process, and data from cyber-physical systems are impacted by the actions/decisions made by experts/engineers to control/stabilize a system. The dynamics, complexity, scale and uncertainty of many practical systems or phenomena necessitate the maximum extraction of information carried by data for accurate learning/modeling process. The innovation of this project resides in the fact that it enables optimal incorporation of the user information during the learning process, as well as learning from multiple data acquired by non-similar users/experts. This project will advance the state of the art in learning and data gathering processes, and contribute to the science base of machine learning, control/learning theory and Bayesian statistics. Original contributions are expected in: 1) Informative Bayesian Learning through Experts’ Data for allowing maximum extraction of information from experts’ data for dynamics or policy learning; 2) Multi-Fidelity Bayesian Optimization framework for efficient learning of very large systems with possibly huge amount of uncertainty; 3) Multiple-Expert Bayesian Learning for efficient incorporation of multiple experts’ data during the learning process; and 4) Near Optimal Bayesian Deep Reinforcement Learning Data Gathering for acquiring the most informative data. The applicability to practical systems will be the key feature of the proposed frameworks in this project due to their rigorous statistically-oriented nature that enables risk-based, efficient, real-time and scalable learning and decision making.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.