Reinforcement Learning-based Lightpath Establishment in All-Optical WDM Networks

Izumi Koyanagi (0851046)


In these days, WDM networks are developed in connecting each node effectively. In this thesis, I propose a new method for utilizing each node by Q-learning, which is one of the reinforcement learning techniques, and when a new lightpath establishment request arrives, the node decides whether the lightpath request should be accepted or rejected according to the most effective action. Then, the node learns the optimal action based on the current state and the used action. After the adequate learning, the node can provide the service differentiation and the effective wavelength utilization. This method can be available when the number of wavelengths and the number of classes are large without the assumptions about the lightpath arrival process and the distribution of the lightpath holding time. Therefore, this method is effective to utilize in real environment. In addition, I implement the reinforcement-based lightpath establishment with Generalized Multi-Protocol Label Switching (GMPLS). In this implementation, at first, I modify Path message, which is used in Resource reSerVation Protocol (RSVP), so that the message supports the information about service class. Then, I implement a reinforcement learning component into each optical switch. I also modify somewhat the processing of some messages. I evaluate the performance of the proposed method with simulation. Furthermore, from numerical examples, I indicate effectiveness of our proposed method in real environment.