[Paper Review] A Deep Generative Adversarial Architecture for Network-Wide Spatial-Temporal Traffic State Estimation

Yunyi Liang, Zhiyong Cui+3 more|arXiv (Cornell University)|Jan 5, 2018

Traffic Prediction and Management Techniques13 references3 citations

TL;DR

This paper proposes a deep generative adversarial architecture (GAA) that combines traffic flow theory with two Long Short-Term Memory (LSTM) networks to estimate network-wide spatial-temporal traffic states with missing data. The generative network learns to produce realistic traffic matrices, while the discriminative network validates them, enabling high-accuracy imputation of missing flow and density values, outperforming Bayesian networks on real-world data from Seattle and San Diego.

ABSTRACT

This study proposes a deep generative adversarial architecture (GAA) for network-wide spatial-temporal traffic state estimation. The GAA is able to combine traffic flow theory with neural networks and thus improve the accuracy of traffic state estimation. It consists of two Long Short-Term Memory Neural Networks (LSTM NNs) which capture correlation in time and space among traffic flow and traffic density. One of the LSTM NNs, called a discriminative network, aims to maximize the probability of assigning correct labels to both true traffic state matrices (i.e., traffic flow and traffic density within a given spatial-temporal area) and the traffic state matrices generated from the other neural network. The other LSTM NN, called a generative network, aims to generate traffic state matrices which maximize the probability that the discriminative network assigns true labels to them. The two LSTM NNs are trained simultaneously such that the trained generative network can generate traffic matrices similar to those in the training data set. Given a traffic state matrix with missing values, we use back-propagation on three defined loss functions to map the corrupted matrix to a latent space. The mapping vector is then passed through the pre-trained generative network to estimate the missing values of the corrupted matrix. The proposed GAA is compared with the existing Bayesian network approach on loop detector data collected from Seattle, Washington and that collected from San Diego, California. Experimental results indicate that the GAA can achieve higher accuracy in traffic state estimation than the Bayesian network approach.

Motivation & Objective

• To address the challenge of accurate network-wide traffic state estimation when sensor data contains missing values.
• To integrate traffic flow theory with deep learning for improved spatiotemporal modeling of traffic dynamics.
• To develop a generative adversarial framework that learns realistic traffic state patterns from incomplete observations.
• To enhance estimation accuracy beyond traditional Bayesian network approaches in real-world urban networks.

Proposed method

• The GAA employs two LSTM neural networks: a generative network that synthesizes traffic state matrices and a discriminative network that distinguishes real from generated matrices.
• The two networks are trained adversarially to improve the realism and accuracy of generated traffic states.
• A three-loss back-propagation framework maps corrupted traffic matrices with missing values into a latent space for reconstruction.
• The latent vector is then passed through the pre-trained generative network to estimate missing flow and density values.
• The model integrates physical traffic flow principles into the deep learning architecture to improve generalization and plausibility.
• Training is conducted on loop detector data from Seattle and San Diego to capture real-world spatiotemporal patterns.

Experimental results

Research questions

• RQ1 Can a deep generative adversarial architecture effectively estimate missing spatial-temporal traffic states across a network?
• RQ2 How does the integration of traffic flow theory with deep learning improve estimation accuracy?
• RQ3 To what extent does the GAA outperform Bayesian networks in traffic state estimation on real-world data?
• RQ4 Can the adversarial training of two LSTMs generate realistic traffic matrices that reflect real-world dynamics?

Key findings

• The GAA achieved higher estimation accuracy than the Bayesian network approach on both Seattle and San Diego loop detector datasets.
• The model effectively captured complex spatiotemporal correlations in traffic flow and density across multiple sensors.
• The adversarial training framework successfully improved the realism and fidelity of generated traffic matrices.
• The three-loss back-propagation method enabled effective mapping of incomplete traffic matrices to a meaningful latent space for accurate imputation.