by Theresa Pepin
People say you can’t predict the future. But the fact of the matter is that we must, or we can’t keep up with the present. When disaster strikes or chaos threatens, or when we need to get emergency personnel in while we’re getting evacuees out, we need to know a great deal ahead of “real” time in order to make decisions quickly enough.
Coming into the second year of his JDRD project, Lee Han and his team have built the foundation for a faster-than-real-time traffic simulation system some 1,000 times faster than the state-of-the-practice.
Melding a lot of simultaneous and heterogeneous data to develop and validate robust algorithms at scale—together with the learning, as Han says, that comes from where “the rubber meets the road”—allows for projection of likely scenarios for a range of possible decisions. Doing this very quickly provides managers and decision makers with readily available, solid evidence in good time.
In traffic simulation, it is essential to preserve the sequential nature of the data. We all know what this means because we have intimate knowledge at an individual or “microscopic” level of driver-reaction times and behavior in stop-and-go driving and traffic jams. We all can also understand what a successful, overall system promises: safety; security in times of terrorist attacks, mass evacuations, and major accidents; fuel savings; and emission reduction.
Massive heterogeneous real-time traffic data is required to build the foundation for a faster-than-real-time, random traffic simulation system. In addition to being 1,000 times faster than the present state-of-the-practice, likely scenarios must be simulated at least thirty to fifty times each for validation. Soon, highway and on-board wireless sensors will add even more to the data mix and flood.
Clearly, the scale and challenges of the required fusion of data and calculations—and the “sequential” constraint in developing parallel algorithms for petascale computing—is enormous.
Han has a large and diverse group of gifted students involved in this project, including PhD candidate Ryan Overton; post-doc Qiang Yang; and undergraduate students Allyson Foster, Clement Oigbokie, and Michael Raup. Two doctoral students—Jianjiang Yang, with extensive background in transportation engineering, and Wei Lu, with training and expertise in computer science—have worked with the computational scientists of the corresponding LDRD project led by James Horey of ORNL’s Computational Sciences and Engineering Division. Together, they have designed a new distributed programming model for optimizing traffic flow and have tested it in Han’s large-scale microscopic traffic simulation system.
At least two publications are in the works on a field theory proposed by Han and his students and UTSim (a Universal Traffic Simulation algorithm based on the field theory) with follow-up funding proposals in process to the Department of Transportation, Department of Energy, and National Science Foundation. Han has recently lectured and met with colleagues on the associated computational challenges of traffic simulation in Berlin and Sarajevo. In order to extend and generalize the research, his students have also traveled on UT’s McClure Travel Scholarship to collect driving data in different countries.
Distributed computation framework for faster-than-real-time microscopic traffic simulation (Year 2)
Lee Han, UT Department of Civil and Environmental Engineering
Distributed Computational Framework for Massive Heterogeneous Data Fusion
James Horey, ORNL Computational Sciences and Engineering Division