A MULTI-AGENT REINFORCEMENT-LEARNING SYSTEM FOR DYNAMIC TRAFFIC MANAGEMENT IN SOFTWARE-DEFINED NETWORKS

A multi-agent routing system for software-defined networks (SDN) is studied for the case where several TCP flows arrive almost simultaneously. The system combines six Q-learning agents, an initial minimum-hop routing policy, and a decision coordinator. It is implemented in Mininet with a Ryu controller on a six-switch topology. Its performance is compared with variants without learning and without coordination, a coordinated greedy heuristic, Dijkstra’s algorithm with fixed link weights, and adaptive AWR routing. The main and supplementary series comprise 40 independent training runs and 1080 tests. Q-learning has no significant effect on throughput: almost all of the gain comes from the system without learning, and the contribution of coordination is not established with sufficient precision. The system outperforms Dijkstra’s algorithm when two flows compete for shared links, by +8.1 Mbit/s for a shared receiver and +4.2 Mbit/s for cross flows. Its advantage over AWR is observed only for the shared receiver, and its advantage over the greedy heuristic exceeds 1 Mbit/s only for cross flows. Reproduction on a different testbed confirmed the advantage for the shared receiver. Conclusions are limited to one topology and two simultaneous flows.