大规模分布式演化算法及其在云计算资源管理中的应用

Evolutionary algorithms have been widely recognized as a class of powerful optimization algorithms. However, as their computational efficiency is quite low and they do not explicitly consider the fault-tolerance of the algorithms, they can hardly be applied into real-world optimization problems. In recent years, the research of distributed evolutionary algorithms has improved the computational efficiency and the fault-tolerance of evolutionary algorithms to some extent, while it still cannot effectively solve many large-scale real-time applications. This project aims to solve the large-scale real-time optimization problems by studying the large-scale distributed evolutionary algorithms. Firstly, we study how to further improve the computational efficiency and the fault-tolerance of existing distributed evolutionary algorithms by paralleling them on a more fine-grained level. And then, for the difficulty that existing distributed evolutionary algorithms cannot solve large-scale problems well, we study how to reduce the solution space effectively so that the original large-scale solution space can be transformed into low-dimensional sub-problems, which can be solve more effectively by existing distributed evolutionary algorithms. Next, these two techniques are optimally integrated to form the targeted large-scale distributed evolutionary algorithm that equips with both high computational efficiency and high fault-tolerance. Lastly, the proposed large-scale distributed evolutionary algorithm is customized to solve several realistic large-scale real-time problems in the area of cloud resource management.

演化算法已被广泛认可为一种问题求解能力较强的优化算法。然而，由于其计算效率较低，且并未显式考虑容错性问题，通常难以被应用到实际优化问题中。近年来，对分布式演化算法的研究已可以在一定程度上提升其计算效率和容错性，但还不能有效解决现实生活中广泛存在的大规模实时性优化问题。本项目拟面向该类问题，研究大规模分布式演化算法。首先，通过更细粒度地按个体对演化算法进行并行化，来进一步提升现有分布式演化算法的计算效率和容错性；然后，针对现有分布式演化算法无法有效求解大规模问题的难点，研究分布式解空间约减技术，将原问题高效转换为低维子问题，以供现有分布式演化算法求解；接着，从分布式技术的角度出发，对前述两种技术进行有机地融合，得到具有高计算效率和高容错性的大规模分布式演化算法；最后，针对云计算资源管理中的若干大规模实时优化问题，对上述得到的大规模分布式演化算法进行定制化设计并实际应用求解。

演化算法已被广泛认可为一种问题求解能力较强的优化算法。然而，由于其计算效率较低，且并未显式考虑容错性问题，通常难以被应用到实际优化问题中。本项目针对上述问题，从演化算法的种群和变量两个维度研究大规模分布式演化算法，提出了具有清晰数学描述的负相关并发探索机制以及求解性能无损的并行协同演化框架，形成了可用于大规模分布式运行的高效演化算法新理论、新框架与新方法。同时，针对云原生系统中多资源关联型弹性伸缩问题，通过对上述方法进行定制化设计形成了高效的解决方案，该方案已成功应用于深交所云行情数据系统中。. 项目组共发表SCI检索的期刊论文8篇（含SCI一区期刊论文3篇），EI检索的会议论文5篇（含CCF-A类会议Spotlight论文1篇）；申请专利15项，授权2项；培养硕士研究生9人，顺利毕业5人；项目组整体超额完成项目任务（发表论文4-6篇，申请专利1-2项，培养研究生2-3名）。项目负责人杨鹏3次在国际国内学术会议上作特邀报告，入围中国人工智能学会优博。基于所研方法与技术，项目负责人杨鹏首先在2021年完成了与深圳证券交易的法定信息披露全资子公司深圳证券信息有限公司合作的“基于云原生技术的高性能信息系统的关键技术研究”项目，进而作为负责人在2022年建立了“深圳证券信息有限公司-南方科技大学计算机系行情云联合创新实验室”。