基于多目标混合蚁狮优化的算法选择方法  

Algorithm Selection Method Based on Multi-Objective Hybrid Ant Lion Optimizer

作　　者：李庚松 刘艺 郑奇斌 李翔 刘坤 秦伟 王强 杨长虹 Li Gengsong;Liu Yi;Zheng Qibin;Li Xiang;Liu Kun;Qin Wei;Wang Qiang;Yang Changhong(National Innovation Institute of Defense Technology,Beijing 100071;Academy of Military Sciences,Beijing 100091;Tianjin Artificial Intelligence Innovation Center,Tianjin 300457)

机构地区：[1]国防科技创新研究院,北京100071 [2]军事科学院,北京100091 [3]天津(滨海)人工智能创新中心,天津300457

出　　处：《计算机研究与发展》2023年第7期1533-1550,共18页Journal of Computer Research and Development

基　　金：科技部科技创新2030-重大项目(2020AAA0104802);国家自然科学基金项目(91948303);国家自然科学基金青年科学基金项目(61802426)。

摘　　要：算法选择是指从可行算法中为给定问题选择满足需求的算法,基于元学习的算法选择是应用较为广泛的方法,元特征和元算法是其中的关键内容,而现有研究难以充分利用元特征的互补性和元算法的多样性,不利于进一步提升方法性能.为了解决上述问题,提出基于多目标混合蚁狮优化的算法选择方法(SAMO),设计算法选择模型,以集成元算法的准确性和多样性作为优化目标,引入元特征选择和选择性集成,同时选择元特征和异构元算法以构建集成元算法;提出多目标混合蚁狮算法对模型进行优化,使用离散型编码选择元特征子集,通过连续型编码构建集成元算法,应用增强游走策略和偏好精英选择机制提升寻优性能.使用260个数据集、150种元特征和9种候选算法构建分类算法选择问题来进行测试,分析方法的参数敏感性,将多目标混合蚁狮算法与4种演化算法进行比较,通过对8种对比方法与所提方法进行对比实验,结果验证了所提方法的有效性和优越性.Algorithm selection refers to selecting an algorithm that satisfies the requirements for a given problem from feasible algorithms,and algorithm selection based on meta-learning is a widely used method,in which the key components are meta-features and meta-learners.However,existing research is difficult to make full use of the complementarity of meta-features and the diversity of meta-learners,which are not conducive to further improving the method performance.To solve the above problems,a selective ensemble algorithm selection method based on multiobjective hybrid ant lion optimizer(SAMO)is proposed.It designs an algorithm selection model,which sets the accuracy and diversity of the ensemble meta-learners as the optimization objectives,introduces meta-feature selection and selective ensemble,and chooses meta-features and heterogeneous meta-learners simultaneously to construct ensemble meta-learners;it proposes a multi-objective hybrid ant lion optimizer to optimize the model,which uses discrete code to select meta-feature subsets and constructs ensemble meta-learners by continuous code,and applies the enhanced walk strategy and the preference elite selection mechanism to improve the optimization performance.We utilize 260 datasets,150 meta-features,and 9 candidate algorithms to construct classification algorithm selection problems and conduct test experiments,and the parameter sensitivity of the method is analyzed,the multi-objective hybrid ant lion optimizer is compared with four evolutionary algorithms,8 comparative methods are compared with the proposed method,and the results verify the effectiveness and superiority of the method.

关 键 词：算法选择 多目标蚁狮优化 元特征选择 选择性集成 元学习 分类 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]