Dynamic Resource Allocation in Integrated Terrestrial-Satellite Networks: An Elastic Decision Transformer Approach

The integration of terrestrial and satellite networks offers broader coverage and higher capacity, meeting the rising demand for connectivity. This paper presents a novel resource allocation approach for integrated terrestrial-satellite networks by utilizing an Elastic Decision Transformer (EDT) to optimize user association and power control. EDT, based on the Decision Transformer (DT), which incorporates trajectory stitching and dynamic history length adjustment, can effectively handle dynamic and complex environments. Different from traditional reinforcement learning methods, which require retraining in new scenarios with varying input numbers, EDT leverages the transformer's ability to process variable-length inputs with ensuring adaptability and fast convergence. Simulation results that the proposed EDT-based approach shows robust convergence, superior generalization capabilities, and scalability across various network scenarios, highlighting its effectiveness in optimizing resource allocation in next-generation communication networks.