Reinforcement Learning-Driven GNSS Observation Selection For Enhanced PPP Accuracy

This paper presents a reinforcement learning (RL) framework integrated into the GMV GSharp® Precise Point Positioning (PPP) algorithm to optimize Global Navigation Satellite System (GNSS) measurement processing. The RL agent, originally developed to reduce multipath effects, has evolved into a decision-making tool that evaluates the usefulness of GNSS measurements, improving positioning accuracy. This change was necessary to address various challenges in GNSS measurement processing, including multipath interference, measurement noise, and other factors that affect PPP performance.

The RL model processes GNSS observations on an epoch-by-epoch basis, utilizing a set of features derived from GNSS measurements, such as pseudoranges, signal-to-noise ratios, elevation angles, and residuals. These features are critical as they provide the necessary data for the RL agent to assess the quality and reliability of each measurement, thereby informing its decision-making process. We designed the action space as a binary decision framework, where the agent determines whether each measurement should be included in the solution. We designed a reward function to encourage decisions that reduce positioning errors. It assigns higher rewards for actions that improve accuracy and penalizes those that undermine the solution's stability.

Training data includes more than 50 hours of GNSS raw observations that we collect across diverse environments, such as urban canyons, suburban areas, and open spaces. This diversity ensures that the RL agent is exposed to multiple real-world conditions, crucial for developing a robust model capable of handling challenges in GNSS measurement processing. Broad validation demonstrates that the RL-enhanced PPP algorithm achieves significant accuracy improvements compared to the baseline GSharp® solution, even under adverse conditions.

By going beyond multipath mitigation, this work shows the potential of RL to evolve the processing of GNSS measurements. For instance, the RL-enhanced PPP algorithm has shown improvements in positioning accuracy and robustness, as evidenced by validation results. The agent’s ability to adaptively manage observation quality and relevance offers a robust and precise positioning solution, facilitating next-generation GNSS applications for complex environments.