Constructing Diagnostic Explanations Using Schema-Structured Bayesian Networks

本研究的目标是开发表征诊断推理的表示和算法。问题领域的人类专家经常在特定因果假设的背景下解释该领域的数据。这种类型的推理与演绎推理形成对比，演绎推理是从一组一般规则和事实中，通过合理的推理规则推断出进一步的信息。然而，诊断推理，通常被称为溯因推理，是从一组事实转向对这些事实存在的“最佳解释”。这个过程通常需要专家做出一个假设性的推测来解释事实，然后寻找具体的新信息来证实这个推测。因此，人类专家在观察到的信息的可能解释空间中搜索。本研究利用贝叶斯信念网络建立一个领域的因果模型。这种方法以一种精确的方式表达了因果模式之间的相互关系，以及它们在解释过程中的应用。该研究支持对相互矛盾和模棱两可的证据的处理，以及对合理推论进行评级的明确方法，以及从现有统计数据中学习条件概率相对优势的可能性。虽然该研究领域是建立在对离散元件半导体故障的研究以及复杂实时控制中故障机制的分析的基础上，但诊断推理是一个普遍的研究领域。研究结果对医疗决策建模、集成电路故障分析以及过程实时监测和控制具有重要意义。http://www.cs.unm.edu/CS_Dept/faculty/homepage/luge r /

The goal of this research is to develop a representation and algorithms that characterize diagnostic reasoning. Human experts in a problem domain frequently interpret data in that domain in the context of a particular causal hypothesis. This type reasoning contrasts with deductive inference where from a set of general rules and facts further information is deduced by sound inference rules. Diagnostic reasoning, often called abductive inference, however, moves from a set of facts to the "best explanation" for the existence of these facts. This process often requires the expert to make a hypothetical conjecture that would explain the facts and then search for specific new information that can confirm that conjecture. Thus the human expert searches through a space of possible explanations for the observed information. This research uses Bayesian Belief Networks to build causal models of a domain. This approach represents in a precise way the interrelationship of causal patterns and their use in moving towards an explanation. The research supports handling of conflicting and ambiguous evidence, as well as a clear method for rating plausible inferences and the possibility of learning relative strengths of conditional probabilities from available statistical data. Although the research domain is built on data from investigation of failures of discrete component semiconductors as well as the analysis of failure mechanisms in complex real time control, diagnostic reasoning is a general research area. Results could be important in modeling medical decision-making, integrated circuit fault analysis, as well as used in real time process monitoring and control. http://www.cs.unm.edu/CS_Dept/faculty/homepage/luge r/