Abstract

Gibbs samplings is a Markov Chain Monte Carlo technique for estimating conditional probabilities in Bayesian networks. A major problem of Gibbs sampling is the dependency of the generated chain of samples. Thus the estimates are biased unless the initial value of the chain is drawn from the target distribution. One elegant method to overcome the initial bias is regenerative samplings. We reported elsewhere the “stationary minorization condition” that makes any Markov Chain Monte Carlo technique regenerative. In this paper, we show how this condition can be easily met in the simulations of any Bayesian network.

中文翻译：

---

贝叶斯网络：再生吉布斯采样

摘要

吉布斯采样是一种马尔可夫链蒙特卡罗技术，用于估计贝叶斯网络中的条件概率。吉布斯采样的一个主要问题是生成的样本链的依赖性。因此，除非链的初始值是从目标分布中提取的，否则估计是有偏差的。一种克服初始偏差的优雅方法是再生采样。我们在别处报告了使任何马尔可夫链蒙特卡洛技术再生的“固定小化条件”。在本文中，我们展示了如何在任何贝叶斯网络的模拟中轻松满足此条件。