Variational Inference for Agent-Based Models with Applications to Achieve Fuel Economy

Wen, Dong; Chunming, Qiao · 2017 · ROSA P / Transportation Informatics University Transportation Center

archive: archived pipeline: cataloged verified

Get this paper ↗ (full text — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of tracking and predicting real-time traffic dynamics in city-scale transportation networks using sparse data from mobile phones and probe vehicles. While the Internet of Things provides abundant vehicle location data, existing methods struggle to estimate traffic states at unobserved times and locations or to fuse these noisy sensor inputs with transportation models. The authors aim to bridge agent-based modeling and machine learning to create a computational platform that combines simulation modeling with big data. This integration allows for the estimation of traffic conditions, such as flow and congestion, and supports applications like interactive driving planners that optimize for fuel economy and travel time. The methodology identifies an agent-based transportation simulator as a stochastic process, specifically a Markov process induced by multi-agent interactions. The system state comprises vehicle locations and trip plans alongside road segment states (vehicle count and congestion status). To handle the exploding state space inherent in tracking thousands of links and vehicles, the authors employ variational inference with mean field approximation. This approach minimizes the Bethe variational principle, allowing the probabilistic evolution of individual links and vehicles to be determined by the average effects of others, rather than searching the joint probability space. The framework utilizes a forward-backward schema to infer hidden states from observations of probe vehicles, treating the simulator as a discrete-time state-space model. Additionally, the authors formulate a multi-agent discrete event decision process to model decentralized decision-making under incomplete information. The proposed variational inference (VI) algorithm was evaluated against deep neural networks (DNN), recurrent neural networks (RNN), and extended Kalman filters (EKF) using three datasets: a synthetic network (SynthTown), a semi-realistic Berlin dataset with 24,000 links, and a real-world Dakar dataset with 8,000 links. Performance was measured using the coefficient of determination ($R^2$) and mean squared error (MSE). Results demonstrated that VI outperformed all other methods, achieving the lowest MSE and highest $R^2$ across tracking and prediction tasks. VI’s superiority is attributed to its ability to explicitly leverage problem-specific structures, such as road topology, and its capacity to handle arbitrary probability distributions, unlike the Gaussian assumptions of EKF. Furthermore, VI exhibited better scalability, as DNN, RNN, and EKF failed to process the larger Dakar dataset. The significance of this work lies in demonstrating that combining simulation modeling with big data via variational inference can accurately track and predict city-scale traffic dynamics. The findings suggest that model-driven approaches that explicitly incorporate structural constraints outperform purely data-driven deep learning methods in this domain. This framework enables the delivery of real-time travel information to drivers, promoting efficient driving and fuel economy by associating specific agent trips with travel times and consumption. The study highlights the potential for integrating machine learning with agent-based models to solve complex inference problems in transportation networks where observations are sparse and noisy.

Key finding

Variational inference outperformed deep neural networks, recurrent neural networks, and extended Kalman filters in tracking and predicting city-scale traffic dynamics, achieving the lowest mean squared error and highest coefficient of determination across SynthTown, Berlin, and Dakar datasets.

Methodology

modeling

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via bulk_ingest_rosap on 2026-05-23 (6 acquisition events logged).

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• exposure measurement

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Theoretical Contribution: computational model