基于高斯混合模型的轨迹模仿学习表征参数优化  

Parameter Optimization of Trajectory Imitation Learning Characterization Based on Gaussian Mixture Model

作　　者：于建均[1,2] 郑逸加 阮晓钢[1,2] 赵少琼[1,2] 

机构地区：[1]北京工业大学电子信息与控制工程学院,北京100124 [2]计算智能与智能系统北京市重点实验室,北京100124

出　　处：《北京工业大学学报》2017年第5期719-728,共10页Journal of Beijing University of Technology

基　　金：国家自然科学基金资助项目(61375086);高等学校博士学科点专项科研基金资助项目(20101103110007);北京市自然科学基金资助项目/北京市教育委员会科技计划重点资助项目(KZ201610005010)

摘　　要：针对高斯混合模型(Gaussian mixture model,GMM)参数选取效率较低的问题,提出了一种在基于GMM的轨迹模仿学习表征中综合求解GMM参数估计的方法.该方法基于多中心聚类算法中的最大最小距离算法改进kmeans算法,得到最优初始聚类中心,并基于贝叶斯信息准则(Bayesian information criterion,BIC)通过遗传算法优化求解,同时获取GMM的4个重要参数.该方法通过提高划分初始数据集的效率,在优化初始聚类中心基础上确定混合模型个数,有效地避免了因为初值敏感而导致的局部极值问题.通过多组仿真实验验证了该方法的有效性.To tackle the problem of low efficiency of parameter selection for Gaussian mixture model （GMM）, an integrated method was proposed to estimate the parameters of GMM in characterization of trajectory imitation learning in this paper. The k-means algorithm was improved by this mothod based on maximum minimum distance algorithm which belonged to multi-center clustering algorithm and the optimal initial cluster center was obtained. Four significant GMM parameters were obtained at the same time by using genetic algorithm based on the Bayesian information criterion （BIC）. The efficiency of dividing the initial data set was im proved, and the initial cluster center and the number of the mixture model was determined, and the local extreme problem caused by initial value was effectively avoided. Multiple sets simulation results demonstrate the validity of this method.

关 键 词：高斯混合模型(GMM) 遗传算法 模仿学习 贝叶斯信息准则(BIC) 最大最小距离算法 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]