贝叶斯网络推理的一种仿真算法

贝叶斯网络是一种强有力的不确定性知识表达和推理工具。网络的推理是贝叶斯网络的重要内容之一。该文提出了一种近似仿真算法。由随机数发生器产生随机数，并按节点的先验概率，由赌轮对网络各个节点状态赋值，得到一个采样样本序列。当样本序列的数量足够大时，边缘统计量和条件统计量与节点的边缘概率和条件概率接近，从而得到网络的近似推理结果。仿真结果表明，该算法与精确解接近，有较好的适应性。基于该算法构造的贝叶斯诊断网络系统已成功应用于天津石化炼油厂。

A Simulation Algorithm for Bayesian Network Inference

Bayesian Network is a powerful tool used to express and infer uncertain knowledge. The network inference is its important content. This paper proposes an approximate algorithm, in which the stochastic sampling process is based on a random number generator and a roulette wheel. The roulette wheel determines the node states according to its prior probabilities. When the amount of sample serials is enough, margin and condition statistic quantity are close to margin and condition probability of node respectively. Then approximate inference results of network are obtained. Numerical simulating results show its effectiveness. The algorithm has been used for Bayesian Diagnostic Network in Tianjin Petroleum- Chemical Complex.