Adaptive Parameter Tuning for Fused Localization Model Using Multi-Agent Reinforcement Learning

doi:10.1109/tvt.2026.3679732

Adaptive Parameter Tuning for Fused Localization Model Using Multi-Agent Reinforcement Learning

2026-03-31

0

PRE

AI

DOI：10.1109/tvt.2026.3679732

Original

Original request for help

Save

Abstract

En 中文

In fully automatic operation of railways, it is taken as a promising solution of high-precise localization to fuse ultra wide band (UWB) and inertial measurement unit (IMU). However, the inherent errors caused by non line of sight (NLoS) propagation and IMU cumulative errors are inevitable, which always leads to similar localization errors along the same track. Different from existing localization schemes by spatio-temporal correlations from one train, we propose data mining scheme among different trains from the view point of multi-agent reinforcement learning (RL). Firstly, UWB ranging data filtering mechanism is modeled as a deep Q-network (DQN), which can kick out ranging data under serious NLoS interferences. Secondly, we propose a parameter tuning scheme for fused localization model to update adaptively, which integrates data among multi-train along the same track. Finally, semi-supervised rewards are evaluated to facilitate DQN training and minimize cumulative localization errors, with the help of unit point locators. Simulation results show that proposed scheme outperforms other three typical fused schemes, with gains of 11.59%, 9.74%, 4.43% in LoS scenarios and 22.13%, 13.77%, 14.77% in NLoS scenarios.

Keywords:

Data-model driven

RL-integrated

fused localization

spatio-temporal mining

Adaptive Parameter Tuning for Fused Localization Model Using Multi-Agent Reinforcement Learning

Abstract

Journal

Organization