Evolutionary Reinforcement Learning for Multi-objective SFC Deployment

In the 5G era, the diversification of application scenarios and the expansion of service demands necessitate a paradigm shift in mobile communication development to ensure Quality of Service (QoS) for a multitude of network services, thereby enhancing user experience and optimizing network resource management. To address this, we extend the traditional single-objective optimization model of Service Function Chains (SFC) to a multi-objective optimization model, incorporating considerations such as latency, deployment cost, and throughput. However, traditional evolutionary computation struggles to optimize multiple objectives simultaneously without considering initial population quality, while Deep Reinforcement Learning (DRL) faces challenges in determining weights between multiple objectives, requiring repeated training and optimization. In this paper, we propose a two-stage solution multi-objective evolutionary reinforcement learning (MOERL) to deploy SFC. In the first stage, DRL can effectively generate excellent initial population solutions, while in the second stage, using this solution as the initial solution for NSGA-II can obtain the required placement solution and reduce computational time. Extensive experiments demonstrate that MOERL significantly reduces computational time, with its efficacy increasing as the initial population size grows. Furthermore, MOERL exhibits superior performance across three objective functions under varying request numbers, outperforming DRL by 20.9% and NSGA-II by 38.2% in computational time reduction. Thus, MOERL adeptly meets the rigorous resource and real-time demands within5G network environments.