Dynamic Packet Routing Algorithm Based on Multidimensional Information and Multiagent Reinforcement Learning

ABSTRACT Packet routing is one of the critical factors that affect network performance and security, with the goal of finding the optimal path for network packets from the source node to the destination node. However, with the diversification of network architectures, the differences in network application requirements, and the time‐varying characteristics of network topologies, the limitations of traditional model‐ and rule‐based routing algorithms in terms of computational overhead and flexibility are becoming increasingly apparent. This paper designs a packet routing strategy based on multiagent deep reinforcement learning (MIMRL). In MIMRL, each router node is abstracted as an independent agent with its own neural network. Multidimensional data such as the current location of the data packet, the number of nodes in the network, the length of the data packet received at the current location node, and the set of neighboring nodes are used as inputs to the neural network. Combined with a segmented reward function, the optimal routing action is determined. Experimental results under different network loads in static and dynamic networks show that the MIMRL algorithm significantly outperforms the benchmark algorithm in multiple metrics such as average delivery time and proportion of full capacity nodes.