Learning to Move Cities: Deep Meta-Models and Reinforcement Policies for Calibration and Control in Urban Networks
Shared latent traffic models calibrate urban simulators and cut travel time by up to 51% with deep Q-learning.
The paper links urban traffic-simulator calibration and reinforcement-learning control through a shared latent representation. A combinatorial MLP-autoencoder maps origin-destination demand and network parameters to travel times and congestion, supporting sample-efficient Bayesian calibration. A deep Q-learning agent with experience replay and target networks then adjusts scheduling and routing. On benchmark networks, the method reduces system-wide travel times by up to 51% compared with baseline operations.
- MLP-autoencoder learns low-dimensional traffic manifolds for Bayesian calibration
- Deep Q-learning adjusts scheduling and routing on compressed traffic state
- Reported system travel times fall by up to 51% versus baseline
From the original article
1 min read at arxiv.orgUrban transportation networks present complex optimization challenges spanning calibration of high-fidelity simulators and real-time operational control. This paper presents a shared latent-space framework that connects simulator calibration and reinforcement learning control through a common learned representation of urban traffic dynamics.
By Adewumi Augustine Adepitan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.