Multi-agent deep reinforcement learning algorithm with self-adaption division strategy for VNF-SC deployment in SDN/NFV-Enabled Networks.

Network function virtualization can decouple the traditional network function from the dedicated hardware, abstracts the software-based virtual network function from the specialized network equipment, and promotes the fundamental transformation of network service deployment mode. However, the deployment of virtual network function (VNF) service chain is an important and crucial problem and key technology faced and must be rescued. In this paper, the problem of VNF service chain deployment in SDN/NFV-Enabled Networks is investigated. The existing solution strategies based on optimization methods (dynamic programming, linear programming, etc.) and heuristic methods (genetic algorithm, particle swarm optimization, etc.) are only suitable for operation deployment in the case of predictable operations, and it is difficult to meet the real-time support operation scheduling requirements in high dynamic combat scenarios. A new real-time algorithm for VNF service chain deployment based on multi-agent deep reinforcement learning with self-adaption division strategy (MDRL-SaDS) to minimize energy consumption in a period of time is proposed. In proposed algorithm, an oriented self-adaptive strategy to determination the number of agents and the optimal division method of VNF service chain for the markov process modeling is designed. Constructing a new neural network model and design a training strategy of joint supervised and unsupervised learning. The global and long-term benefits are used to optimize the scheduling process, and the decision-making framework of offline learning and online deployment is used to solve the VNF service chain deployment problem. Finally, experimental results indicate that the MDRL-SaDS has more advantages and has higher convergence speed, average reward value and stability than compared algorithms, while decreasing the energy consumption in a period of time.