Multipath Parameter Estimation Based on Reinforcement Learning

Multipath as a major error source for Global Navigation Satellite System (GNSS) navigation and positioning in urban environments is difficult to eliminate. Existing works show that supervised learning-based technologies have achieved good multipath mitigation performance for simulation data. However, due to the unavailability of precisely labeled real data, training the model using simulation data will cause performance degradation for real data due to data mismatch problem. To address this problem, this paper employs reinforcement learning (RL) to mitigate the multipath effect through multipath parameter estimation. Unlike supervised learning, reinforcement learning agent learns by interacting with the environment instead of pre-training. The RL-based estimator uses maximum likelihood estimation (MLE) criterion with multiple sampling points of the autocorrelation function (ACF) as the input. The simulation results show that the performance of the RL-based estimator is affected by the sampling spacing setting of ACFs, and the smaller the sampling space, the better the performance. The RL-based estimator achieves good short multipath mitigation performance when the ACF is sampled at a spacing of 0.025 chips, in which case, the multipath-induced range error is reduced to less than 3.2 meters.