CAREER: Explanation, Decision Making, and Learning in Graphical Models

图形模型，如贝叶斯网络和影响图，为解决不确定问题下的推理和决策提供了原则性的方法。然而，现有方法对这些图形模型的适应性和可扩展性往往是有限的。这个项目旨在通过开发新的和改进的方法来解释、决策和在图形模型中学习，来解决其中的一些限制。它包括以下具体目标：(1)开发新的方法来找到贝叶斯网络中仅包含给定观测的最相关变量的解释，(2)开发基于启发式搜索的方法和算法来更有效地求解影响图，(3)开发新的算法来学习特定领域的启发式信息指导的最优贝叶斯网络，使得只需要探索一小部分解空间，以及(4)将本项目中开发的方法应用于现实世界，包括多故障诊断、供应链风险管理和在线协作学习。这个项目可以在许多学科的不确定性下产生更好的推理和决策方法，这些学科的图形模型已经得到了成功的应用，包括医学、安全、规划、商业、经济、教育等。该项目还可以导致开发新的和改进的课程和课程，让代表人数不足的群体的学生参与研究，并通过免费软件、出版物和演示文稿广泛传播研究成果。

Graphical models, such as Bayesian networks and influence diagrams, provide principled approaches to solving reasoning and decision making under uncertainty problems. However, the adaptability and scalability of existing methods for these graphical models are often limited. This project aims to address some of these limitations by developing new and improved approaches to explanation, decision making, and learning in graphical models. It includes the following specific objectives: (1) developing new approaches to finding explanations that only contain the most relevant variables for given observations in Bayesian networks, (2) developing heuristic search-based methods and algorithms to solve influence diagrams more efficiently, (3) developing new algorithms for learning optimal Bayesian networks guided by domain-specific heuristic information so that only a small fraction of the solution space need to be explored, and (4) applying the methods developed in this project to real-world applications including multiple-fault diagnosis, supply chain risk management, and online collaborative learning. This project can lead to significantly better approaches to reasoning and decision making under uncertainty in many disciplines where graphical models have found successful applications, including medicine, security, planning, business, economics, education, and many others. This project can also lead to the development of new and enhanced courses and curricula, the involvement of students from underrepresented groups in the research, and a wide dissemination of the research outcomes through free software, publications, and presentations.