Graph-guided MADQN based handover strategy for LEO satellite networks

Abstract

In low earth orbit satellite networks, the high orbital velocity of satellites causes frequent communication link handovers, which adversely affects user service quality and the efficiency of network resource management. To maintain communication continuity, an effective handover strategy is required. Traditional handover strategies are typically based on fixed decision rules, which makes it difficult to balance handover frequency, signal quality, and network load balancing. Reinforcement learning based methods, while capable of adapting to dynamic network conditions, suffer from low training efficiency due to the absence of prior knowledge, and are also susceptible to unstable convergence when applied to high dimensional state spaces. To address these issues, this article proposes a graph-guided multi-agent deep Q-network handover strategy. First, by leveraging the predictability of satellite orbits, we precompute the shortest handover path from the current time to the end of the communication session using a directed graph. Then, we encode this path as prior guidance information into the agents’ state space in two forms: immediate recommendations and future planning information over multiple steps. This enables the agents to combine current network states with the future path guidance from the graph model when making decisions. Simulation results show that our proposed strategy performs better than the baseline schemes in handover count, throughput, and load balancing. This confirms the effectiveness of combining guidance from the graph model with multi-agent reinforcement learning.