Abstract: Cloud computing is widely applied across various fields,and the workflow scheduling problem has attracted significant attention.To address the issues of diverse user demands and instability in computational resource capabilities,the workflow scheduling problem was treated as a dynamic multi-objective problem,and a dynamic adaptive cascading clustering and reference point incremental learning algorithm was proposed.By considering task completion time,cost,load,average resource utilization rate and reliability five indicators,a dynamic multi-objective workflow scheduling model was constructed.To prevent the algorithm from falling into local optima due to dynamic changes in the scheduling environment over time,the proposed algorithm introduced a restart mechanism and designed reinforcement learning for adaptive parameter selection to improve the convergence and solution diversity of the algorithm.Experiments conducted on the FDA series of test problems with known Pareto fronts and on four well-known real-world workflows,the result demonstrated that the proposed algorithm had significant advantages.

Key words: cloud computing, dynamic many-objective optimization, workflow scheduling, reinforcement learning, evolutionary algorithm

摘要： 云计算广泛应用于各个领域,其工作流调度问题备受关注。针对不同用户需求以及计算资源计算能力不稳定问题,将工作流调度问题视为动态多目标问题,并提出一种动态自适应级联聚类和参考点增量学习算法。首先,考虑任务完工时间、成本、负载、平均资源利用率和可靠性5个指标,构建动态多目标工作流调度模型;然后,针对调度环境随时间动态变化使得算法陷入局部最优的问题,所提算法引入重启机制,并设计强化学习进行自适应参数选择,以提高算法的收敛性和解的多样性。在帕累托已知的FDA系列测试问题以及4种著名的真实工作流上进行实验,结果表明本文提出的算法有显著优势。

关键词: 云计算, 动态多目标优化, 工作流调度, 强化学习, 进化算法