An Enhanced Reptile Algorithm for Massive MIMO CSI Prediction: Learning for Robust Adaptation
Abstract
Accurate channel state information (CSI) is crucial in massive multiple-input multiple-output (MIMO) systems, motivating the need for reliable CSI prediction. Although deep learning-based methods show strong potential, their adaptability to environmental variations is constrained. Meta-learning methods, notably Model-Agnostic Meta-Learning (MAML), have been introduced to enhance adaptability. However, they pose significant hurdles for practical deployment, primarily because MAML’s performance heavily relies on fine-grained and sensitive hyperparameter tuning. This paper first introduces a Reptile-based method for CSI prediction to enable robust adaptation in new environments. To further improve robustness and adaptation capability, we introduce three tailored enhancements to the vanilla Reptile. First, we propose a consistent task generation strategy to ensure intra-task consistency. Second, we design a multi-task aggregation strategy to ensure inter-task diversity. Third, we introduce a momentum inner optimizer to stabilize the optimization trajectory. The results indicate that our method adapts more robustly and rapidly using a few new samples across a wide range of settings and generally outperforms other benchmarks, underscoring its practical deployment value.