Twin Adaptive Distributional Reinforcement Learning for Robust Handover Control in LEO Satellite Networks

The inherent high mobility of low Earth orbit (LEO) satellite networks creates dynamic topologies and complex resource scenarios that challenge robust handover control, causing existing RL methods to face challenges including temporal-difference (TD) target noise, which can lead to value overestimation and decision instability. This letter proposed a twin adaptive distributional reinforcement learning (TAD-RL) algorithm for robust handover control in LEO satellite networks. TAD-RL leverages twin distributional critics to mitigate value overestimation, incorporates expectation-value substitution to suppress TD target noise, and applies variance-based adaptive gradient scaling for stable, accelerated convergence under said heterogeneous reward scales. Simulation results demonstrate that TAD-RL achieves up to 25% higher throughput and a 20% reduction in handover frequency compared to baseline approaches, while maintaining stable performance under heavy network loads.