A Deep Reinforcement Learning Based Dynamic Resource Allocation Approach in Satellite Systems

Abstract

Efficient resource allocation in space information networks (SINs) is crucial for providing global connectivity but is challenged by constrained satellite resources and dynamic user demand. While dynamic channel allocation techniques exist, they often fail to handle complex, multi-faceted resource constraints in practical scenarios. To address this issue, this paper introduces a deep reinforcement learning based dynamic resource allocation (DDRA) algorithm. The DDRA formulates the allocation problem as a Markov decision process and employs deep Q-network (DQN) to learn an optimal policy for assigning channel, power, and traffic resources. We developed a simulation environment in ns-3 to evaluate the DDRA algorithm against traditional fixed and greedy random allocation methods. The results demonstrate that the DDRA algorithm significantly outperforms these baselines, achieving substantially lower service blocking rates and higher traffic satisfaction rates across various user demand scenarios. This work validates the potential of DRL to create intelligent, adaptive resource management systems for next-generation satellite networks.