Planning of Primary and Alternative Path Using Deep Learning Model for Traffic Engineering in Small Networks

Traffic engineering enables to control of network congestion and achieves load balancing using paths planning method for traffic flows. However, optimum path planning requires complex calculations, and it requires huge calculation time. We have been proposing a path planning method using the deep learning model reducing computational complexity and attaining load balancing. In this method congestion-controlled paths are designed; those paths are while looking over the entire network in response to traffic demand consisting of traffic between multiple locations. In this paper, we focus on alternative path design. An alternative path is designed separate from the primary-path. And the alternative path is used when a network node accident or a link failure occurs. This path is designed depending on a purpose, such as node-disjoint-paths or link-disjoint-paths. The path excludes any nodes or links which are contained in the primary-path. Particularly, alternative paths should be designed according to the primary-path. Therefore, even when path planning using deep learning, it is desirable that a single learning model simultaneously designs the primary-path and alternative-path. However, in the path planning methods using deep learning, learning models, which learn different teaching data and simultaneously, design different paths have not been considered. Thus, we extend the conventional learning model. The model outputs two paths for one input. As teaching data, the primary-path is created using the Dijkstra method, and the alternative-path is created as a link-disjoint-path considering the primary-path. The experimental result shows that both paths are designed with a very high success rate of path planning and are confirmed superior in terms of low link overlapping rate.