贝叶斯网络推理的一种高精度仿真算法

High Precision Simulation Algorithm for Bayesian Network Inference

简要分析了贝叶斯网络推理算法的现状,提出了基于马氏链随机拟蒙特卡罗算法(MCRQMC)的推理方法。在给出高精度推理结果的同时,该推理算法亦能给出推理结果的标准偏差。从理论上对MCRQMC算法与现有的算法进行了比较分析,并采用随机Halton序列、Sobol序列和普通随机序列进行了推理实验。结果表明MCRQMC算法在同样样本数量的情况下,推理精度显著优于现有算法。

The current research status of inference methods for Bayesian Networks was reviewed, and a new inference algorithm based on Markov Chain Randomized Quasi-Monte Carlo （MCRQMC） was proposed. The new algorithm could provide high precision inference result and corresponding standard error at the same time. Comparative analysis of MCRQMC and currently used algorithms was conducted theoretically and experimentally. The experiments on the randomized halton sequence, sobol sequence and ordinary randomized sequence demonstrate that MCRQMC outperforms the conventional algorithms in terms of inference precision.