Toward Packet Routing With Fully Distributed Multiagent Deep Reinforcement Learning

Packet routing is one of the fundamental problems in computer networks which a router determines next-hop each packet queue to get it as quickly possible its destination. Reinforcement learning (RL) has been introduced design autonomous policies with local information stochastic arrival and service. However, curse dimensionality RL prohibits more comprehensive representation dynamic network states, thus limiting potential benefit. In this article, we propose novel framework based on multiagent deep (DRL) possess an independent long short term memory (LSTM) recurrent neural (RNN) for training decision making fully distributed environment. The LSTM RNN extracts features from rich regarding backlogged packets past actions, effectively approximates value function Q-learning. We further allow route communicate periodically direct neighbors so that broader view state can be incorporated. experimental results manifest our multiagent DRL policy strike delicate balance between congestion-aware shortest routes, significantly reduce delivery time general topologies compared counterparts.