Machine learning techniques to improve router configurations in large optical networks

点击翻译按钮获取中文摘要

We propose to use machine learning techniques to automate the tuning of router configurations in large optical networks; automated tuning should lead to better responsiveness to congestion issues and less error-prone operation. Through our proposed collaboration with Ciena we will have access to log data that captures Interior Gateway Protocol (IGP) metrics such as inbound and outbound traffic on links in a large network, as well as configuration changes performed by human operators. Using supervised learning, we propose to learn in which context the configuration changes are performed. Next, we will create a simulation that is capable of roughly generating the same data as the logs we have, where we will apply the model that we learned in the previous stage. This will give us a baseline performance evaluation. To improve on our initial model, we propose to use reinforcement learning in the simulator obtained in the previous step. Finally, we propose to test the resulting reinforcement learning policy in a small lab with actual routers; this should helps us evaluate what changes may be necessary and which risks to expect when going "live".