A Meta-Reinforcement Learning Method for Enhancing GNSS Positioning by Integrating Spatiotemporal Correlations

Global navigation satellite system (GNSS) positioning is highly susceptible to non-line-of-sight and multipath effects, which significantly limit positioning accuracy. To address these issues, this paper proposes a novel meta-reinforcement-learning-based positioning method that incorporates spatiotemporal correlation and dynamically adapts to the GNSS positioning environment. Key parameters from the satellite receiver are extracted to construct environmental features. These features are processed using convolutional neural networks and long short-term memory networks and integrated with proximal policy optimization (PPO) to build a model termed the CL-PPO model. This base model is further enhanced with a model-agnostic meta-learning algorithm to form the CL-MPPO model, enabling rapid adaptation to new environments. In comparative experiments on Google’s public data sets, CL-MPPO demon-strates superior convergence and generalization, outperforming all five existing methods. CL-MPPO improves the Decimeter Challenge score by 63.19% and reduces the average error to 27.66% compared with raw GNSS measurements, while sastifying real-time positioning requirements.