Traffic in cities generally follows an intermittent pattern linked to the typical 9-to-5 business hours. In any case, this periodic pattern gets disturbed following an unfortunate accident. And currently, the major challenge that traffic engineers face is designing accurate traffic flow models, for use during such mishaps, that they must adapt to during unanticipated traffic situations in real-time. To overcome this issue, a team of researchers from the Lawrence Berkley National Lab is working in collaboration with the California Department of Transportation (Caltrans) to use machine learning (ML) and high-performance computing (HPC) that can help improve Caltrans’ decision-making capabilities in real-time, whenever accidents take place.
The researchers conducted this study together with California Partners for Advanced Transportation Technology (PATH), part of UC Berkeley’s Institute for Transportation Studies (ITS), and Connected Corridors, a collaborative program to study, build and test an Integrated Corridor Management approach to take control of the current traffic scenarios in California.
The system is currently being implemented by Caltrans and Connected Corridors on a trial basis in Los Angeles County through the I-210 pilot. By gathering data from reps in Southern California on the city, county, and state levels, the objective of this research is to improve Caltrans’ real-time decision-making capabilities. This can be done by implementing synchronized multi-jurisdictional traffic incident response plans to reduce the number of accidents happening lately. The first batch of this system is planned to be deployed in the cities of Pasadena, Arcadia, Monrovia, and Duarte in 2020, with future deployment plans across the state.
“Currently, several approaches related to traffic-flow prediction exist, and each can be beneficial in the right circumstance,” said Sherry Li, a mathematician in Berkeley Lab’s Computational Research Division (CRD). “To erase the disbelief of depending on human operators who blindly trust one specific model, our objective was to include numerous models that provide firmer and accurate traffic predictions. We did this by structuring an ensemble learning algorithm that integrates other sub-models.
The concept of ensemble learning has been around for a while and machine learning researchers have explored its capabilities for a long time. It is an ML paradigm that combines an assorted set of learners (singular models) to improve, on the fly, the steadiness, and predictive power of the model. What is uncommon about traffic flow is the transient characteristic; traffic flow estimations are correlated after some time, similar to the prediction results from various models.
In this collaborative research, the ensemble model considers the mutual reliance of sub-models and assigns the “shares of vote” to balance their individual performance with their co-dependency. The ensemble model likewise values recent prediction performance more than older historical performance. Ultimately, the model is superior to any of the individual models used in testing in both stability and prediction accuracy.
The research began with funding from Berkeley Lab’s Laboratory Directed Research and Development (LDRD) program. The objective was to develop a computational system that would enable HPC applications explicit to transportation, for example, streamlining and control of traffic flow. Brian Peterson, a systems development manager at PATH, leads the entire team and also manages the systems development team at Connected Corridors. Hongyuan Zhan, a former Berkeley Lab Computing Sciences summer student from Penn State, was a major contributor to the Connected Corridors work for this research study.