Deep-learning-based path computation without routing convergence in optical satellite networks

Low Earth orbit (LEO) satellite networks, which are composed of multiple inter-connected satellites, have become important infrastructure for future communications. Benefiting from the high bandwidth and anti-interference of satellite laser communication, optical satellite networks, in which satellite links are lasers, can provide global Internet services and have become a research trend. The orbit at a lower altitude has advantages such as low latency, low cost, and easy deployment in LEO optical satellite networks. Meanwhile, the movement of satellites is fast and thus will result in frequent changes for ground–satellite links. The conventional static routing strategy cannot perceive the network state; therefore, the static routing is inapplicable in the case of link failure or congestion. Dynamic routing can ensure the accuracy of the network connection by routing convergence. However, the routing table needs to be updated frequently because of the highly dynamic topology, resulting in the increase in signaling overhead. To compute routing paths accurately while reducing the update frequency of the routing table, this paper proposes a path computation model based on deep learning. By learning the mapping relation of previous services and the routing paths, the model can directly output the routing path according to the current service request. Using this method, the path computation tasks depend less on the frequently updated routing table. The simulation results show that the paths computed by the proposed method are almost the same as the paths computed by Dijkstra’s algorithm, the average accuracy rate is above 90%, and the highest accuracy rate can reach 98.8%. Compared with traditional path computation, the proposed method needs to collect a large amount of previous data for training, and the training time is about several hours.