Improving the Traffic Engineering of SDN networks by using Local Multi-Agent Deep Reinforcement Learning

The flexibility and programmability of Software-Defined Networks (SDN) has allowed the research community to propose new Traffic Engineering (TE) techniques to improve their performance. Although the installation of heuristic or optimal solutions in the SDN controller allows to obtain good results in network performance, these are based on historical data that may not be updated to actual traffic variations. Moreover, the research community is exploiting the strength of Deep Reinforcement Learning (DRL) and its capability to learn and adapt to the complexities inherent in networks to propose enhanced routing solutions. However, the nature of DRL can cause a performance degradation during the learning process due to the application of exploration when determining the best policy. For this reason, in this work we propose a Multi-Agent DRL (MADRL) based solution that is able to reduce the Maximum Link Utilization (MLU) of SDN networks only considering the local information of the nodes. Each node has a DRL agent and is able to decide the best routing decision for each flow so that the MLU is minimized. The performance evaluation shows that our approach outperforms the classical shortest path rule based on Dijkstra in 8%.