Gradient-Based Task-Aware Meta-Learning for Fingerprint-Based Localization in Cell-Free Massive MIMO Systems

The deployment of cell-free massive multiple-input–multiple-output (CF-mMIMO) systems in urban areas enriches wireless channel characteristics, making it a promising approach for fingerprint-based localization. However, the high cost of label collection and the unreliable generalization performance significantly limit its widespread application. Recently, model-agnostic meta-learning (MAML) has achieved remarkable success in few-shot learning by extracting common knowledge from existing tasks. But the use of a forcibly shared meta-parameter for model initialization often struggles with task heterogeneity in practical applications. To address these challenges, we propose a novel gradient-based task-aware meta-learning (GTML) framework for fingerprint-based localization. We first model the localization problem in a new environment with limited fingerprint data as a meta-learning problem for new task adaptation. Then, we propose an improved embedded task-aware method based on training gradients to reduce the overhead of task-specific feature extraction. The proposed GTML includes two paradigms using task-specific information to customize the global meta-learner: for a few historical tasks, a weighted paradigm is introduced to compensate for task heterogeneity; for a larger set of historical tasks, a clustered paradigm is used to capture the distribution of training tasks and learn group-specific meta-parameter. The simulation results using wireless insite (WI) software demonstrate that the proposed GTML enables rapid adaptation to a new environment with a few training samples. Moreover, compared to the vanilla MAML, GTML improves the localization accuracy by over 11% while maintaining low computational overhead.