| Abstract |
This work presents a new routing framework based on ma- chine learning that attempts to increase flexibility, resilience and inter-domain privacy. In Neural Routing, a central con- troller continuously trains a deep neural network with the results of a mixed integer linear program minimizing net- work congestion, using samples of flows from the network as inputs. The trained model is periodically sent to switches which can then forward packets without the need for ex- plicit flow rules. This combines the flexibility of SDN with close-to-optimal network utilization. The framework intro- duces resilience in two dimensions: firstly, if the controller is unavailable, the model should operate independently with minimal impact on efficiency for the period until a new con- troller is elected. Secondly, link failures can be directly pro- vided to the model in the switches allowing routing decisions to adapt to the new topology without needing explicit prior decisions by the central controller. Finally, Neural Routing allows for privacy preserving collaboration between routing domains/autonomous systems by exporting a derived neural network model that allows assertions on performance met- rics, capacity and policies in remote domains. |