Modern Bayesian network inference and its practical applicaitons in probabilistic expert systems

人们经常不得不在不确定的情况下做出决定。在计算机系统中，不确定性管理的一种方法是使用贝叶斯网络。为了做出合理的决策，贝叶斯网络使用概率理论作为坚实的数学基础来回答问题(查询)。它们已经被许多公司使用，包括微软、NASA、洛克希德、惠普和诺基亚。要高效、快速地回答问题，关键是要利用问题领域中持有的独立性。然而，目前的方法并没有充分利用所有可用的独立性。通过利用在所有以前的方法中仍然没有注意到的独立性，我们最近提出了第一种方法，可以准确地识别在贝叶斯网络中传递的概率信息。此外，我们识别这些消息的速度大大快于它们的物理构造。我们的方法可以探测贝叶斯网络，以指导回答问题所需的物理计算。要记住的一点是，通过这样做，我们可以比以前更快地回答问题。贝叶斯网络界也普遍认为，初学者很难理解查询处理算法。原因是这些以前的算法在回答问题时无法识别在贝叶斯网络中传递的消息。我们可以的。这项研究计划将产生两个主要成果。首先，我们将建立一个最先进的计算机系统，将贝叶斯网络用于不确定性管理。其次，我们将编写一本关于贝叶斯网络的教科书，描述我们的新方法。由于我们的方法可以更准确地描述推理过程，它应该使更广泛的受众能够访问贝叶斯网络。

One often has to make decisions under uncertainty. In computer systems, one approach to uncertainty management is to use Bayesian networks. In order to make sound decisions, Bayesian networks use probability theory as a solid mathematical foundation to answer questions (queries). They have been used by many companies, including Microsoft, NASA, Lockheed, Hewlett Packard and Nokia. To answer questions efficiently and quickly, the key is to exploit independencies holding in the problem domain. Current approaches, however, are not fully taking advantage of all available independencies. By utilizing independencies that remain unnoticed in all previous methods, we have recently suggested the first method that can precisely identify the probability information being passed in a Bayesian network. In addition, we can identify these messages significantly faster than they can be physically constructed. Our method can scout the Bayesian network in order to guide the physical computation needed to answer a question. The point to remember is that by doing so we can answer questions faster than before. It is also widely acknowledged in the Bayesian network community that beginners have great difficulty understanding query processing algorithms. The reason is that these previous algorithms are unable to identify the messages being passed in a Bayesian network when answering a question. We can. This research program will produce two major results. First, we will build a state-of-the-art computer system for using Bayesian networks in uncertainty management. Second, we will write a textbook on Bayesian networks describing our new approach. Since our approach can more precisely describe the inference process, it should make Bayesian networks accessible to a wider audience.