深度强化学习优化的学习向量量化聚类算法  

Learning vector quantization clustering algorithm for deep learning optimization

作　　者：史凯岳 李凤莲 张雪英 杜海文 于丽君 SHI Kaiyue;LI Fenglian;ZHANG Xueying;DU Haiwen;YU Lijun(School of Information and Computer Science,Taiyuan University of Technology,Jinzhong 030600,China;Shanxi CETC New Energy Technology Co.,Ltd.,Taiyuan 030024,China)

机构地区：[1]太原理工大学信息与计算机学院,山西晋中030600 [2]山西省中电科新能源技术有限公司,山西太原030024

出　　处：《电子设计工程》2023年第9期43-48,共6页Electronic Design Engineering

基　　金：国家自然科学基金资助项目(62171307);山西省科技重大专项(20181102008)。

摘　　要：基于学习向量量化(Learning Vector Quantization,LVQ)单次迭代聚类效果不稳定和随着数据维度增大,聚类效果下降的缺陷,采用了一种深度强化学习优化的LVQ聚类算法。将LVQ算法的每一次迭代看做深度强化学习的一个状态,LVQ算法初始化一组原型向量后,用原型向量与数据集中每一个数据点做“拉近”或“远离”运算来完成一次迭代。优化算法挑选一部分数据点,并与原型向量做运算,将这一过程作为一个动作,选取的数据子集不同,产生的动作也不同,把这些动作组成动作集,选定动作后,根据奖赏函数找到最佳动作,进入下一状态。通过对UCI公共数据集和碳碳沉积数据集试验,得出优化后的算法FMI提升3%到10%,Dunn指数提升2%到9%,准确率提高3%到6%,用于公共数据集及碳碳沉积材料的生产过程数据分析性能较优。Based on the defects of Learning Vector Quantization(LVQ)that the clustering effect is unstable in a single iteration and the clustering effect decreases with the increase of data dimension,a LVQ clustering algorithm optimized by deep reinforcement learning is adopted.Each iteration of LVQ algorithm is regarded as a state of deep reinforcement learning.After initializing a set of prototype vectors,LVQ algorithm completes an iteration by"drawing in"or"moving away"from each data point in the data set.The optimization algorithm selects a part of data points,and operates with the prototype vector,and regards this process as an action.Different data subsets select different actions,and these actions are grouped into action sets.After selecting actions,the best action is found according to the reward function,and the next state is entered.Experiments on UCI public data set and carbon deposition data set show that the FMI is improved by 3%to 10%,the Dunn index is improved by 2%to 9%,and the accuracy rate is improved by 3%to 6%,the data analysis performance of public data set and carbon deposition material production process is better.

关 键 词：深度Q网络 学习向量量化 聚类 碳碳沉积 

分 类 号：TN98[电子电信—信息与通信工程]