A Flexible Topology Reconstruction Strategy Based on Deep Q-Learning for Balance Performance and Efficiency of STINs

Abstract

Due to the mobility and vulnerability of satellite nodes, the topology of satellite-terrestrial integrated networks (STINs) is highly dynamic. It requires a flexible reconstruction strategy to avoid severe performance loss. However, the current reconstruction strategies usually result in a heavy computational burden, especially in the case of a shortage of satellite resources. Hence, it is of great practical interest to design a flexible topology reconstruction strategy to balance the computational efficiency and network performance of STINs. We investigated this important problem in this paper and made the following contributions. First, a deep Q-learning model is proposed to dynamically optimize the node classification results and flexibly determine the range of network reconstruction, which could enhance the efficiency of the network recovery. Second, a modified SVM classification model is proposed to classify the node type with the hyperplane parameters dynamically adjusted, which can increase the accuracy of the classification result and improve the convergence efficiency of the reconstruction algorithm. Third, an artificial bee colony algorithm is proposed to flexible recover the performance of the damaged network, which considers the age of information and range of the topology cross iteration. Numerical results demonstrate that compared with existing algorithms, the average end-to-end delay decreased by about 13.68%, and other performance indicators of STINs where improved to a certain extent.