Teaching the Travel Demand Flow Estimation Models: A New Deep-Learning Approach Using Multi-Source Data

Abstract
The amount of data that transportation systems collect and store today has reached a new level compared with the previously traditional collection methods. For example, as one of pioneering Metropolitan Planning Organizations (MPO) in the United States in transportation planning, Maricopa Association of Governments (MAG) can provide the regional household travel survey data of 60GB, 1-year TMC-based speed data of 26GB, and 1-year link-based speed data of 3.1TB. Facing those unstructured and structured data streaming in from heterogeneous sensor sources at an unprecedented rate, it is critical to quickly manage and mine useful information under control. Also, it should be always aware that the value of those big data is reflected not just by its high volume but also by what specific goals/problems the data are used for. With the development of new computing technologies, machine learning has currently evolved as a powerful tool to learn from data with independent adaptions to generate reliable, repeatable decisions and results in a variety of application areas. However, compared with the areas in automotive, financial services, healthcare and etc., the open-source learning materials, studies and applications of machine learning on transportation system planning and operations are still relatively weak and need to be enhanced and taught to students, researchers and practitioners in time.

Focusing on the Traffic Demand Flow Estimation (TDFE) problem, which infers the number of persons/vehicles traveling between a particular origin and destination via a particular route/link, this project aims to support the education and training of relative students, researchers and practitioners to understand and learn the knowledge of deep learning and its application procedure by developing a step-by-step tutorial and open-source software packages and providing a number of well-organized workshops.

Specifically, deep learning technologies have been applied in a number of studies mainly focusing on traffic flow prediction (Dougherty, 1995; Park et al., 1998; Dia, 2001; Yin et al., 2002; Vlahogianni et al., 2005; Zhong et al., 2005; Zheng et al., 2006; Chan et al., 2012; Kumar et al. 2013). Recently, Lv et al. (2015) demonstrated one of the early applications of deep learning networks to uncover and identify hidden patterns from the observed traffic measurements as a time series from multiple days. It has been well recognized that the simple application of the artificial neural network (ANN) software package is insufficient in explaining the behavioral relationship between complex traffic states and underlying traveling choice parameters. In this project, our major focus is to develop a theoretically explainable and interpretable deep learning approach to estimate different layers of demand variables and behavior coefficients, and then teach and educate interested users for the learning and real-world applications.

Research Products and Implementation

Scope of Work

​​Final Report (coming soon)