Beijing University of Posts and Telecommunications - multi-agent deep reinforcement learning for cost and delay sensitive virtual network function placement and routing

【 title 】Multi-Agent Deep Reinforcement Learning for Cost- and Delay-Sensitive Virtual Network Function Placement and Routing

【 The author team 】Shaoyang Wang, Chau Yuen, Wei Ni

【 Date of publication 】2022.6.24

【 Thesis link 】https://arxiv.org/pdf/2206.12146.pdf

【 Recommended reasons 】 This paper presents an effective and novel multi-agent based deep reinforcement learning (MADRL) Methods , Used to solve the joint virtual network function (VNF) Layout and routing (P&R), It contains multiple service requests with different requirements at the same time . Differentiated requirements for service requests are reflected in their delays and cost sensitivities . The author first constructs a VNF P&R Problem to jointly minimize the weighted sum of service delay and resource consumption cost , This is a NP complete . then , union VNF P&R The problem is decoupled into two iterative subtasks ： Place subtasks and routing subtasks . Each subtask consists of multiple concurrent sequential decision processes . By invoking the depth deterministic strategy gradient method and multi-agent technology , Designed a MADRL-P&R Framework to perform these two subtasks . New joint reward and internal reward mechanisms are proposed to match the goals and constraints of layout and routing subtasks . This paper also proposes a model retraining method based on parameter migration to deal with the changing network topology .