张力悦1, 2,
苏昭玉1, 2,
张赟3,
张磊3
1.燕山大学信息科学与工程学院 秦皇岛 066004
2.河北省特种光纤与光纤传感重点实验室 秦皇岛 066004
3.北京市机电研究院 北京 100027
基金项目:国家重点研发项目(2019YFB1707301)，河北省人才工程培养资助项目(A201903005)
详细信息
作者简介:刘浩然：男，1980年生，教授，研究方向为贝叶斯算法、工业故障诊断及预测
张力悦：男，1994年生，博士生，研究方向为贝叶斯算法、工业故障诊断及预测
苏昭玉：女，1994年生，硕士生，研究方向为贝叶斯算法、工业故障诊断及预测
张赟：女，1979年生，博士，研究方向为机械设计及原理、系统建模
张磊：男，1991年生，学士，研究方向为数控机床在线测量及系统建模
通讯作者:刘浩然 liu.haoran@ysu.edu.cn
中图分类号:TN911.7计量
文章访问数:253
HTML全文浏览量:177
PDF下载量:54
被引次数:0
出版历程
收稿日期:2020-05-15
修回日期:2021-03-19
网络出版日期:2021-04-15
刊出日期:2021-07-10
Bayesian Variational Inference Algorithm Based on Expectation-Maximization and Simulated Annealing
Haoran LIU1, 2,,,Liyue ZHANG1, 2,
Zhaoyu SU1, 2,
Yun ZHANG3,
Lei ZHANG3
1. School of Information Science and Engineering, Yanshan University, Qinhuangdao 066004, China
2. The Key Laboratory for Special Fiber and Fiber Sensor of Hebei Province, Yanshan University, Qinhuangdao 066004, China
3. Beijing Institute of Mechanical and Electrical Engineering, Beijing 100027, China
Funds:The National Key Research and Development Program of China (2019YFB1707301), Hebei Talent Engineering Training Support Project(A201903005)
摘要
摘要:针对贝叶斯变分推理收敛精度低和搜索过程中易陷入局部最优的问题，该文基于模拟退火理论(SA)和最大期望理论(EM)，考虑变分推理过程中初始先验对最终结果的影响和变分自由能的优化效率问题，构建了双重EM模型学习变分参数的初始先验，以降低初始先验的敏感性，同时构建逆温度参数改进变分自由能函数，使变分自由能在优化过程得到有效控制，并提出一种基于最大期望模拟退火的贝叶斯变分推理算法。该文使用收敛性准则理论分析算法的收敛性，利用所提算法对一个混合高斯分布实例进行实验仿真，实验结果表明该算法具有较优的收敛结果。
关键词:贝叶斯变分推理/
模拟退火/
最大期望/
逆温度参数
Abstract:For the problem that Bayesian variational inference with low convergence precision is easy to fall into local optimum during search process, a Bayesian variational inference algorithm based on Expectation-Maximization (EM) and Simulated Annealing (SA) is proposed. The influence of the initial prior on the final result and the optimization efficiency of the variational free energy in the process of variational inference can not be ignored. The double EM is introduced to construct the initial prior of the variational parameter to reduce the sensitivity of the initial prior. And the inverse temperature parameter is introducted to improve the free energy function, which makes the energy be effectively controlled in the optimization process. This paper uses convergence criterion theory to analyze the convergence of the algorithm. The proposed algorithm is used for experiments with an Gaussian mixture model and the experimental results show that the proposed algorithm has better convergence results.
Key words:Bayesian variational inference/
Simulated Annealing(SA)/
Expectation-Maximization(EM)/
Inverse temperature parameter
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=e62a9d5c-ecef-4b6f-b1da-5243029f4b4c
