.png){kind=icon}
.png){kind=icon}
Return
Data evacuation optimization using multi-objective reinforcement learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.jnca.2025.104390.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
After some disaster occurs, rapid data evacuation among cloud data centers is of great importance. Data evacuation optimization is a two-stage process including destination selection and flow scheduling. These two stages are related to each other, while evacuation efficiency is affected by evacuation distance, bandwidth allocation ratio, and total amount of evacuation flow at the same time. The mutual constraints among multiple factors make it difficult to find or approximate the optimal solution via single-objective optimization. This paper proposes a new two-stage data evacuation strategy using multi-objective reinforcement learning, with evacuation flow optimization as the central optimization objective across both stages. In the first stage, it simultaneously minimizes total path length and maximizes the total available bandwidth to determine source–destination pair for every evacuation transfer. In the second stage, it simultaneously allocates proportional bandwidth and maximizes the total amount of evacuation flow to find path and allocate bandwidth for every evacuation transfer. Reward function is set based on classifying candidate sets to search for optimal solution while ensuring that feasible solutions are obtained. Chebyshev scalarization function is used to evaluate action rewards and optimize action selection process. Performance comparison is implemented with state-of-the-art algorithms based on different data volumes and network scales. Simulation result demonstrates that the new strategy outperforms other algorithms with higher evacuation efficiency, good convergence and robustness.
Journal
IF:
8
Papers:
3.6K
Citations:
1.1W