1: \begin{abstract}

2: 		The existing indoor fingerprinting-based localization methods are rather accurate after intensive offline calibrations for a specific environment, and they are built based either on the received signal strength (RSS) or the channel state information (CSI). However, a well-calibrated localization method (which can be a pure statistical signal processing method or an emerging data-driven method) will present poor generalization abilities in changing environments, which results in large losses in knowledge and human effort. To break the environment-specific localization bottleneck, we propose a novel data-driven fingerprinting-based localization framework empowered by the model-agnostic meta-learning (MAML), named MetaLoc. Specifically, MetaLoc is characterized by its ability to rapidly adapt itself to a new, possibly unseen, environment with very little calibration. The underlying data-driven localization model is a deep neural network, and we

3: 	leverage historical data previously collected from various well-calibrated environments to train an optimal set of meta-parameters as an initialization to the new environments. Furthermore, we develop two MetaLoc paradigms in the proposed MetaLoc based on the different ways of obtaining meta-parameters. The centralized paradigm using vanilla MAML is much easier to implement, while the distributed paradigm incorporates domain shifts into the vanilla MAML to accelerate the convergence speed of the training process. The experimental results obtained for both synthetic- and real datasets demonstrate MetaLoc's strengthes in terms of localization error, robustness and cost-effectiveness compared with various baseline methods. 

4: \end{abstract}

5: