AP-IS:面向多模态数据的智能高效索引选择模型  

AP-IS:Intelligent and Efficient Index Selection Model for Multimodal Data

作　　者：乔少杰 刘晨旭 韩楠[2] 徐康镭 蒋宇河 元昌安 吴涛 袁冠[5] QIAO Shao-Jie;LIU Chen-Xu;HAN Nan;XU Kang-Lei;JIANG Yu-He;YUAN Chang-An;WU Tao;YUAN Guan(School of Software Engineering,Chengdu University of Information Technology,Chengdu 610225;School of Management,Chengdu University of Information Technology,Chengdu 610225;Guangxi Key Laboratory of Human-machine Interaction and Intelligent Decision,Nanning Normal University,Nanning 530100;School of Cyber Security and Information Law,Chongqing University of Posts and Telecommunications,Chongqing 400065;School of Computer Science and Technology,China University of Mining and Technology,Xuzhou 221116)

机构地区：[1]成都信息工程大学软件工程学院,成都610225 [2]成都信息工程大学管理学院,成都610225 [3]南宁师范大学广西人机交互与智能决策重点实验室,南宁530100 [4]重庆邮电大学网络空间安全与信息法学院,重庆400065 [5]中国矿业大学计算机科学与技术学院,徐州221116

出　　处：《自动化学报》2025年第2期457-474,共18页Acta Automatica Sinica

基　　金：国家自然科学基金(62272066);四川省科技计划(2025YFHZ0194,2025ZNSFSC0044);成都市区域科技创新合作项目(2025-YF11-00031-HZ,2025-YF11-00050-HZ);广西人机交互与智能决策重点实验室开放基金项目(GXHIID2207);成都市技术创新研发项目(2024-YF05-01217-SN)资助。

摘　　要：现有的索引选择方法存在诸多局限性.首先,大多数方法考虑场景较为单一,不能针对特定数据模态选择合适的索引结构,进而无法有效应对海量多模态数据;其次,现有方法未考虑索引选择时索引构建的代价,无法有效应对动态的工作负载.针对上述问题,提出一种面向多模态数据的智能高效索引选择模型APE-X DQN(Distributed prioritized experience replay in deep Q-network),称为AP-IS(APE-X DQN for index selection).AP-IS设计了新型索引集编码和SQL语句编码方法,该方法使AP-IS在感知多模态数据的同时兼顾索引结构本身的特性,极大地降低了索引的存储代价.APIS集成新型索引效益评估方法,在优化强化学习奖励机制的同时,监控数据库工作负载的执行状态,保证动态工作负载下AP-IS在时间和空间上的优化效果.在真实多模态数据集上进行大量实验,验证了AP-IS在工作负载的延迟、存储代价和训练效率等方面的性能,结果均明显优于最新索引选择方法.Existing index selection methods have several limitations. Firstly, most methods only consider a singlescenario and cannot select an appropriate index structure for a specific data modal, thus being unable to effectivelycope with massive multimodal data;Secondly, these methods do not take into consideration the cost of constructingindexes when selecting indexes, making them be unable to effectively handle dynamic workloads. Aiming to copewith these issues, an intelligent and efficient index selection model for multimodal data based on the APE-X DQN(distributed prioritized experience replay in deep Q-network) model, called AP-IS (APE-X DQN for indexselection). AP-IS designs a new index set encoding and a new SQL statement encoding method, allowing AP-IS toperceive multimodal data while considering the characteristics of the index structure itself, significantly reducing thestorage cost of indexes. AP-IS integrates a novel index benefit evaluation method. While optimizing the rewardmechanism of reinforcement learning, it can monitor the execution state of database workloads, guaranteeing the effectof time and space optimization of AP-IS under dynamic workloads. Extensive experiments are conducted onreal multimodal datasets to evaluate the performance of AP-IS under workloads including latency, storage cost,training efficiency, and the results of AP-IS significantly outperforms the state-of-the-art index selection methods.

关 键 词：智能数据库 多模态数据 索引选择 强化学习 执行计划 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论]