Considering the defects of conventional optimization methods, a novel optimization algorithm is introduced in this paper. Target space partitioning method is used in this algorithm to solve multi-objective optimizatio...Considering the defects of conventional optimization methods, a novel optimization algorithm is introduced in this paper. Target space partitioning method is used in this algorithm to solve multi-objective optimization problem, thus achieve the coherent solution which can meet the requirements of all target functions, and improve the population's overall evolution level. The algorithm which guarantees diversity preservation and fast convergence to the Pareto set is applied to structural optimization problems. The empirical analysis supports the algorithm and gives an example with program.展开更多
针对现有多目标进化算法在处理复杂Pareto边界问题时收敛性与多样性难以平衡的问题,提出了一种改进的动态关联策略(dynamic association model,DAM)和基于目标空间转换的自适应权值更新策略(objective-oriented transformation strategy...针对现有多目标进化算法在处理复杂Pareto边界问题时收敛性与多样性难以平衡的问题,提出了一种改进的动态关联策略(dynamic association model,DAM)和基于目标空间转换的自适应权值更新策略(objective-oriented transformation strategy,OTS)的高维多目标进化算法MOEA/D-DAM.该算法通过DAM将权重向量与种群个体通过切比雪夫加权法进行关联,实现种群与权向量的动态关联,增强了种群的收敛性与多样性;采用基于目标空间转换的OTS通过对帕累托前沿曲率估计,当其曲率小于1时,将种群在目标空间进行坐标转换;采用目标向量之间的余弦相似度来衡量稀疏度,有效改善了个体在帕累托前沿上分布不均的问题使种群在保持多样性的同时快速收敛.MOEA/D-DAM与其他先进算法MOEA/D-UR、MOEA/D-URAW、MOEA/D-VOV、PeEA以及TS-NSGA-Ⅱ在DLTZ测试问题和WFG测试问题上进行仿真对比实验.结果表明,MOEA/D-DAM在IGD性能指标上分别有43、48、52、49、40个问题表现优于其他算法,最优解占比为54.6%;在HV性能指标上有28、46、42、37、43个问题依次优于其他算法,最优解占比为42.1%.该算法在求解复杂Pareto边界的高维多目标优化问题中表现出强大的竞争力,能够有效平衡收敛性与多样性.展开更多
基金National Natural Science Foundations of China (No. 60970004, No. 60743010)Natural Science Foundation of ShandongProvince, China (No. Z2008G02)
文摘Considering the defects of conventional optimization methods, a novel optimization algorithm is introduced in this paper. Target space partitioning method is used in this algorithm to solve multi-objective optimization problem, thus achieve the coherent solution which can meet the requirements of all target functions, and improve the population's overall evolution level. The algorithm which guarantees diversity preservation and fast convergence to the Pareto set is applied to structural optimization problems. The empirical analysis supports the algorithm and gives an example with program.
文摘针对遥感图像中大纵横比目标因正样本不足而出现的学习不充分问题,提出一种基于形状自适应标签分配的遥感有向目标检测网络(shape-adaptive label assignment for oriented object detection network,SALANet)。首先,引入纵横比敏感系数建立目标几何特征与正样本数量的动态映射关系,缓解传统方法中固定分配规则引发的样本分布不平衡问题;其次,设计自适应标签分配策略,通过对交并比(intersection over union,IoU)进行排名实现高质量正样本选择;最后,提出中心轴先验,将圆形中心先验区扩展为目标中心轴的矩形区域,增强大纵横比目标的几何特征表征能力。在DOTAv1.0和HRSC2016数据集上的对比实验表明,SALANet分别取得0.777 1和0.932 3的平均精度均值(mean average precision,mAP),较基线方法RoI Transformer分别提升8.15%和2.87%。
文摘针对现有多目标进化算法在处理复杂Pareto边界问题时收敛性与多样性难以平衡的问题,提出了一种改进的动态关联策略(dynamic association model,DAM)和基于目标空间转换的自适应权值更新策略(objective-oriented transformation strategy,OTS)的高维多目标进化算法MOEA/D-DAM.该算法通过DAM将权重向量与种群个体通过切比雪夫加权法进行关联,实现种群与权向量的动态关联,增强了种群的收敛性与多样性;采用基于目标空间转换的OTS通过对帕累托前沿曲率估计,当其曲率小于1时,将种群在目标空间进行坐标转换;采用目标向量之间的余弦相似度来衡量稀疏度,有效改善了个体在帕累托前沿上分布不均的问题使种群在保持多样性的同时快速收敛.MOEA/D-DAM与其他先进算法MOEA/D-UR、MOEA/D-URAW、MOEA/D-VOV、PeEA以及TS-NSGA-Ⅱ在DLTZ测试问题和WFG测试问题上进行仿真对比实验.结果表明,MOEA/D-DAM在IGD性能指标上分别有43、48、52、49、40个问题表现优于其他算法,最优解占比为54.6%;在HV性能指标上有28、46、42、37、43个问题依次优于其他算法,最优解占比为42.1%.该算法在求解复杂Pareto边界的高维多目标优化问题中表现出强大的竞争力,能够有效平衡收敛性与多样性.
