Deep reinforcement learning with robust augmented reward sequence prediction for improving GNSS positioning

Data-driven technologies have shown promising potential for improving GNSS positioning, which can analyze observation data to learn the complex hidden characteristics of system models, without rigorous prior assumptions. However, in complex urban areas, the input observation data contain task-irrelevant noisy GNSS measurements arising from stochastic noise, such as signal reflections from tall buildings. Moreover, the problem of data distribution shift between the training and testing phases exists for dynamically changing environments. These problems limit the robustness and generalizability of the data-driven GNSS positioning methods in urban areas. In this paper, a novel deep reinforcement learning (DRL) method is proposed to improve the robustness and generalizability of the data-driven GNSS positioning. Specifically, to address the data distribution shift in dynamically changing environments, the robust Bellman operator (RBO) is employed into the DRL optimization to model the deviations in the data distribution and to enhance generalizability. To improve robustness against task-irrelevant noisy GNSS measurements, the long-term reward sequence prediction (LRSP) is adopted to learn robust representations by extracting task-relevant information from GNSS observations. Therefore, we develop a DRL method with robust augmented reward sequence prediction to correct the rough position solved by model-based methods. Moreover, a novel real-world GNSS positioning dataset is built, containing different scenes in urban areas. Our experiments were conducted on the public dataset Google smartphone decimeter challenge 2022 (GSDC2022) and the built dataset Guangzhou GNSS version 2 (GZGNSS-V2), which demonstrated that the proposed method can outperform model-based and state-of-the-art data-driven methods in terms of generalizability across different environments.