结合近邻传播聚类的世界生成神经网络优化器  

World-based Generation Neural Network Optimizer Combined with Affinity Propagation Clustering

作　　者：陆敏芳 宗伟 陈美涵 杨波[1] 王琳[1] 张波[1] LU Minfang;ZONG Wei;CHEN Meihan;YANG Bo;WANG Lin;ZHANG Bo(School of Information Science and Engineering,University of Jinan,Jinan 250022,Shandong,China)

机构地区：[1]济南大学信息科学与工程学院,山东济南250022

出　　处：《济南大学学报（自然科学版）》2023年第4期449-456,共8页Journal of University of Jinan(Science and Technology)

基　　金：山东省重点研发计划项目(2019GGX101041);济南市高校院所科研带头人工作室项目(2021GXRC077)。

摘　　要：为了解决黑盒问题优化领域中传统优化算法在学习问题结构时存在缺乏样本多样性的问题,设计结合近邻传播聚类的世界生成神经网络优化器。该优化器通过独特的世界模型与采样生成器进行协同学习,完成对问题结构的学习并生成更加多样化的解,以此为基础提出新的优化算法;将所提出的算法与5种有代表性的算法在12个不同特征的优化问题上进行多角度的对比。结果表明,结合近邻传播聚类的世界生成神经网络优化器在不同特征基准问题上平均性能达到最优,准确度平均排名第一,证明了结合近邻传播聚类的世界生成神经网络优化器在学习问题结构的有效性,同时增加了样本的多样性。To address the problem of lack of sample diversity in traditional optimization algorithms for learning problem structuresin field of black-box optimization,a world-generating neural network optimizer combined with affinity propagation clusteringwas designed.The optimizer learned the problem structure and generates more diverse solutions through a unique world model in collaboration with a sample generator.The proposed algorithm was compared with five representative baseline algorithms on 12 different characteristics of the optimization problem from multiple perspectives.The results show that the proposed world-generating neural network optimizer combined with affinity propagation clustering achieves optimality on average for different characteristics of the baseline problem,and the algorithm is more stable.The effectiveness of the world-generating neural network combined with affinity propagation clustering in learning the structure of the problem is demonstrated,while effectively increase the diversity of samples.

关 键 词：近邻传播聚类 世界生成神经网络 黑盒问题 世界模型 神经网络 协同学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]