一种基于多智能体强化学习的流量分配算法

Most of the research on traditional traffic engineering strategies has focused on constructing and solving mathematical models, and its computational complexity is too high. For this reason, an experience-driven traffic distribution algorithm based on multi-agent reinforcement learning is proposed. This algorithm can effectively distribute traffic on pre-computed paths without the need to solve complex mathematical models, thus efficiently and fully utilizing network resources. The algorithm is centrally trained on a software-defined network controller and is executed distributively on access switches or routers after training is completed, while also avoiding frequent interactions with the controller. Experimental results show that, compared with the shortest path and equal-cost multi-path algorithms, the new algorithm effectively reduces the end-to-end delay of the network and increases the network throughput.

传统的流量工程策略的研究大多集中在构建和求解数学模型方面,其计算复杂度过高,为此,提出了一种经验驱动的基于多智能体强化学习的流量分配算法.该算法无需求解复杂数学模型即可在预计算的路径上进行有效的流量分配,从而高效且充分地利用网络资源.算法在软件定义网络控制器上进行集中训练,且在训练完成后在接入交换机或者路由器上分布式执行,同时也避免和控制器的频繁交互.实验结果表明,相对于最短路径和等价多路径算法,新算法有效减少了网络的端到端时延,并且增大了网络吞吐量.