Abstract: In order to solve the problem of lacking effectiveness in the individuals inference and scientific in the overall integration results, which existed in the conditional probability tables in Bayesian network according to experts knowledge inference, the inferred information reduction mechanism is proposed to make judgment objects more intuitive and judgment modes more simple. Then the pair-wised judgment matrix in the Analytic Hierarchy Process（AHP） is used as an extraction means and information carriers of the subjective conditional probability. After that, the experts relative inference method for deriving optimal conditional probability from macro and micro information is constructed, the proposed method on one hand can overcome the problem of lacking effectiveness in individual inferred information caused by experts’ limited concluding ability in the traditional method, on the other hand it is possible to make effective comprehensive integration of experts’ inferred information to ensure scientific of the results. Then the construction process of Bayesian networks is proposed following an inference sequence of “anterior to later”. Finally a data comparison analysis and a case simulation analysis are employed to prove the present method to be scientific, applicable and feasible.

Key words: Bayesian network, relative inference, expert inference, inferred information reduction, conditional probability table

摘要： 为了解决依据专家知识推断贝叶斯网络中条件概率表时存在的个体推断信息缺乏有效性以及整体集成结果缺乏科学性的问题，提出了有利于实现判断对象更直观、判断方式更简便的推断信息约简机理，然后将层次分析法中两两比较与判断矩阵分别作为主观条件概率信息的提取手段与信息载体，构建了能够从宏微观推断信息中提取最优条件概率的专家相对推断方法，该方法一方面可以克服传统方法中因专家推断能力有限的现实约束而造成个体推断信息容易缺乏有效性的问题，另一方面也能够对专家个体推断信息进行有效综合集成、保证决策结果的科学性。在此基础上遵循由前至后的推断顺序提出了贝叶斯网络的构建过程，最后应用数值对比分析和案例模拟分析验证了提出方法的科学有效性和应用可行性。

关键词: 贝叶斯网络, 相对推断, 专家推断, 推断信息约简, 条件概率表