Scalable link cost learning for real-time route guidance systems

•Developed a learning-based framework for RGS using multi-armed bandit (MAB)•Modeled latent factors as link weights, integrating real-time large-scale GPS data.•Achieved an improvement in compliance rates from 64.22 % to 70.87 % in real-world RGS experiments.•Validated scalability through national-s...

Full description

Saved in:
Bibliographic Details
Published in:Transportation research. Part C, Emerging technologies Vol. 178; p. 105244
Main Authors: Kim, Pooreumoe, Yun, Hyunsoo, Kim, Jungmin, Kim, Dong-Kyu
Format: Journal Article
Language:English
Published: Elsevier Ltd 01.09.2025
Subjects:
Link cost learning
Multi-armed bandit
Online learning
Reinforcement learning
Route guidance system
User compliance
Route guidance system
Link cost learning
User compliance
Online learning
Multi-armed bandit
Reinforcement learning
ISSN:0968-090X
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:•Developed a learning-based framework for RGS using multi-armed bandit (MAB)•Modeled latent factors as link weights, integrating real-time large-scale GPS data.•Achieved an improvement in compliance rates from 64.22 % to 70.87 % in real-world RGS experiments.•Validated scalability through national-scale testing by millions of trip samples.•Demonstrated robust integration with heuristic logic in real-world RGS application. Real-time Route Guidance Systems (RGS) are a critical component of Intelligent Transportation Systems (ITS), designed to provide users with optimal routes by responding to real-time traffic data. Central to RGS is the link cost function, which quantifies the generalized cost of traversing road network links. Existing methods primarily rely on travel time and geometry of links, overlooking latent factors that significantly influence user experience. Additionally, processing millions of links in large-scale networks to account for these factors poses substantial computational challenges. As a solution, we employ a scalable, learning-based framework to estimate link costs by learning latent factors from user responses with multi-armed bandit (MAB). The latent factors are modeled as weights representing the general inconvenience of traversing the links. User non-compliance with guided routes serves as key evidence of a discrepancy between RGS’ information and the real-world’s latent factor. Real-time GPS data is used to estimate travel times, which are adjusted by the link weights to compute link costs to provide real-time routes to users. Our methodology was implemented in Kakao Mobility’s real-time RGS, a leading mobility platform in South Korea, and demonstrated across a national-scale network. Online evaluation proved the methodology’s scalability and effectiveness, processing millions of sample trips while enhancing multiple metrics to evaluate the RGS performance. Particularly, for 11.59 % of cases where route guidance differed from the baseline, the compliance rate increased from 64.22 % to 70.87 %. Moreover, our method integrates with custom adjustments commonly applied by RGS operators, ensuring compatibility with existing RGS and empirical applicability.
ISSN:0968-090X
DOI:10.1016/j.trc.2025.105244