RI: Small: Generalized Anytime Probabilistic Inference

概率推理现在经常用于科学和工程的许多领域，在这些领域，它是执行文本分类、社会网络分析、医疗诊断、信息提取、概率规划、视觉和机器人等任务的系统的基础。这个项目旨在开发一个随时和通用的概率推理引擎，其目标是广泛的概率表示，包括经典的命题表示-例如贝叶斯和马尔可夫网络-除了基于一阶逻辑的更具表现力的表示。该项目还将研究复杂类中的概率查询，这些查询在文献中尚未得到太多关注，但可用于研究基于概率推理系统的推理和决策的稳健性。目标推理引擎计划将最新技术应用于精确推理，为近似推理服务，允许它只有在精确推理产生时才放弃近似推理。此外，该引擎计划随着时间的推移平稳地、渐进地改进其近似值。该项目的一个关键重点是在使用最通用的概率表示作为输入的同时实现这些目标，以允许尽可能广泛地采用所开发的推理引擎。我们的预期结果将影响许多领域，允许用户在更大的模型和不同的背景下执行更准确的概率推理。通过科学文章、研究研讨会、会议报告和研究生教学，我们期望所取得的成果得到广泛传播，以最大限度地发挥不同社区所获得的好处。此外，我们计划公开发布嵌入我们结果的软件系统，延续公开发布获奖概率推理软件系统的悠久传统。

Probabilistic reasoning is now routinely used in many fields of science and engineering, where it underlies systems that perform tasks such as text classification, social network analysis, medical diagnosis, information extraction, probabilistic planning, vision and robotics. This project aims to develop an anytime and generalized probabilistic inference engine that targets a wide range of probabilistic representations, including classical, propositional representations --- such as Bayesian and Markov networks --- in addition to more expressive representations based on first order logic. The project will also investigate probabilistic queries in complexity classes that have not received much attention in the literature, yet can be used to study the robustness of inferences and decisions based on probabilistic reasoning systems. The targeted inference engine is planned to put the state of the art in exact inference at the service of approximate inference, allowing it to resign to approximations only when exact inference yields. Moreover, the engine is planned to smoothly and incrementally improve its approximations over time. A key emphasis of the project is to accomplish these objectives while using the most general probabilistic representation as an input, to allow for the widest possible adoption of the developed inference engine. Our anticipated results will impact many fields by allowing users to perform more accurate probabilistic inference, on larger models and in different contexts. Through scientific articles, research seminars, conference presentations, and graduate teaching, we expect the obtained results to be widely disseminated so as to maximize the attained benefits by various communities. Moreover, we plan to publicly release software systems that embed our results, continuing with a long tradition of publicly releasing award-winning software systems for probabilistic reasoning.