Representations, Inference and Learning for Complex Decision Making Under Uncertainty

该建议是关于在存在丰富关系描述的领域中不确定情况下的决策制定，例如根据患者电子健康记录中的详细观察结果做出治疗决策，或者根据对当地地质的描述做出在地理区域勘探的决策。基于统计关系人工智能、概率规划、本体论推理和偏好启发的最新进展，本提案的目的是为不确定性下的推理和决策设计一个连贯的表达框架。******概率模型与关系模型的混合被称为统计关系人工智能。这些模型指定了逻辑结构（就量化的逻辑变量而言）和概率结构（就条件独立性而言）。利用两者的问题，被称为提升推理，最近已经被一些研究小组从本质上解决了精确的无向关系情况。这个建议的一部分是建立在最近的继承者的基础上，建立下一代具有表达性表示、高效推理和鲁棒学习的关系概率系统。我们还将进一步推进精确推理，这也将提供新的方法来进行有效的近似推理。******最近，科学和政府数据的爆炸式增长，使用正式本体定义的词汇表发布。我们将致力于表示、推理和学习与异构数据集和丰富本体互操作的假设。当在不同的抽象和细节级别定义对多个异构数据集进行预测的多个假设时，会出现令人兴奋的挑战。******最终这些模型是用来做决定的。决策的另一部分是效用。我们还将研究与关系模型和本体交互的实用新型的偏好激发，并为决策者所理解。******我们的工作将建立在与现实世界的决策者在地质学、计算可持续性和医学方面的持续合作基础上。这个建议是发展计算基础，使这种应用可行。***********

This proposal is about decision making under uncertainty in domains where there are rich relational descriptions, such as making treatment decisions about medical patients conditioned on the detailed observations in their electronic health records, or making a decision about prospecting in a geographic region conditioned on a description of the local geology. Building on recent progress in statistical relational AI, probabilistic programming, ontological reasoning and preference elicitation, the aim of this proposal is to design a coherent expressive framework for reasoning and decision making under uncertainty.******The mix of probabilistic models with relational models has become known as statistical relational AI. These models specify both the logical structure (in terms of quantified logical variables) and the probabilistic structure (in terms of conditional independence). The problem of exploiting both, called lifted inference has recently been essentially solved for the exact undirected relational case by a number of research teams. Part of this proposal is to build on the recent successors to build the next generation of relational probabilistic systems with expressive representations, efficient inference and robust learning. We will also further advance exact inference, which will also enable new ways to do efficient approximate inference.******Recently there has been an explosion of scientific and government data being published with the vocabulary defined by formal ontologies. We will work on representing, reasoning and learning hypotheses that interoperate with heterogeneous data sets and rich ontologies. There are exciting challenges that arise when defining multiple hypotheses at various levels of abstraction and detail that make predictions on multiple heterogeneous data sets.******Ultimately these models are used to make decisions. The other part of making decisions is utilities. We will also work on preference elicitation for utility models that interact with the relational models and ontologies, and are understandable by decision makers.******Our work will build on ongoing collaboration with real-world decision makers in geology, computational sustainability, and medicine. This proposal is to develop the computational foundations to make such applications feasible.***********