“RLTE: Reinforcement Learning for Traffic-Engineering”, In Resilient Networks and Services - Second International Conference on Autonomous Infrastructure, Management and Security (AIMS 2008), LNCS 5127, 2008.
Dipl.-Inf. Erik Einhorn
Prof. Dr.-Ing. habil. Andreas Mitschele-Thiel
- Date of publication
- Quality of service (QoS) is gaining more and more impor-
tance in today’s networks. We present a fully decentralized and self-organizing approach for QoS routing and Traffic Engineering in connection oriented networks, e.g. MPLS networks. Based on reinforcement learning the algorithm learns the optimal routing policy for incoming connection requests while minimizing the blocking probability. In contrast to other approaches our method does not rely on predefined paths or LSPs and is able to optimize the network utilization in the presence of multiple QoS restrictions like bandwidth and delay. Moreover, no additional signaling overhead is required. Using an adaptive neural vector quantization technique for clustering the state space a considerable speed-up of learning the routing policy can be achieved. In different experiments we are able to show that our approach performs better than classical approaches like Widest Shortest Path routing (WSP).