Programs & Labs
Subhead H2
Esse molestie consequat. vel illum dolore eu feugiat? Est usus legentis in qui facit eorum claritatem.
Summer 2004
Improving the estimation of travel demand
David Levinson
Traffic congestion is an increasingly serious problem for many of the world's urban areas. As a result, estimating travel demand has become a critical priority to manage ever more congested traffic. Success in employing ITS strategies depends not only on the availability of high-quality, real-time information about traffic conditions, but also on prediction models that make it possible for traffic managers to anticipate the response to proposed actions. At the ITS Institute, a research group led by David Levinson, an assistant professor in the civil engineering department, has developed a practical model to describe the interaction between travel demand and traffic flow phenomena. Since most traffic models use an Origin-Destination (O-D) matrix as the basic description of travel demand, it was necessary to generate estimates of O-D matrices. Levinson's research further developed methods to estimate O-D demand from traffic counts for use in simulating traffic in freeway sections and corridors.
A transportation engineering challenge
The process of demand estimation is an attempt to understand and predict the behavioral patterns of individuals—i.e., the choices made on routes and trips. Over the years, many techniques have evolved to estimate travel demands. The reason is that many traffic management methods rely on simulation for development and testing, creating a need for a travel demand estimation model that can be used in these simulation applications. However, traffic simulation is only as good as its input data, and it is impossible to inexpensively measure entry-ramp to exit-ramp flows, which would be particularly useful for testing ramp metering control strategies, Levinson says. Prior research led by Gary Davis (also with the CE department) and supported by the Minnesota Department of Transportation and the University produced a viable method for estimating freeway O-D patterns from loop detector data and led to the development of microscopic traffic simulation. In 1997, Institute researchers developed a laboratory environment for traffic analysis where various roadway design/operational alternatives could be evaluated with state-of-the-art traffic simulators under an integrated database-simulation environment. Existing freeway simulation models with detailed traffic data, collected using machine-vision and loop detection systems, were evaluated, and one macroscopic and one microscopic model were selected. These models have an integrated user interface and can share basic data. According to Levinson, past traffic simulations of freeway sections used traffic counts (on-ramp and off-ramp) and splitting factors (the percent of traffic exiting at a particular off-ramp) to determine demand and calibrate the simulation. While traffic count data are adequate for certain applications, they are ineffective for more complex applications such as freeway/arterial corridors, freeway networks, or even freeway sections that have major changes such as a closed off-ramp. In those cases, O-D data are required. Unfortunately, O-D data are not routinely collected. Application to real-world traffic The research done by Levinson's team developed methods to estimate O-D demand from traffic counts for use in traffic simulation of freeway sections and corridors. The objectives of this research were twofold: to develop and implement a methodology for estimating O-D demand using available data from traffic counts and to apply that method to estimate demand on specific, real-world corridors. The research team went through this process twice using different strategies each time. The first work, prepared as part of the master's thesis of student Satya Muthuswamy, developed an off-line method using a traffic flow model embedded in a search routine. The traffic flow model—the microscopic traffic simulator AIMSUN—is used in many traffic simulation applications in the Twin Cities. The search routine, which aimed to minimize error between estimated and observed traffic counts, was a gradient-based optimization algorithm, MINOS. These two programs were interfaced to estimate an O-D matrix. When posed this way, the problem was highly non-linear and non-smooth, and the optimization routine found multiple local minima but could not guarantee a global minima. The system exhibited "sensitive dependence on initial conditions." However, with a number of starting "seed" matrices, an O-D matrix with a good fit in terms of reproducing traffic counts can be estimated. The results do depend on initial conditions, which is why multiple matrices need to be used. The experiments indicated that mainline counts were the largest source of traffic in the O-D estimation. The quality of the estimates improved as the error (introduced due to the discrepancy between AIMSUN and the real-world process that generates the on-ramp and off-ramp counts) was reduced.
Learning from the first
The second iteration of the research benefited from what was learned in the first. This work, which formed the basis of the master's thesis of student Yao Wu, examined several methods for estimating O-D matrices for freeways using loop detector data. Least-squares-based methods were compared in terms of both off-line and on-line estimation. Simulated data and observed data were used for evaluating the static and recursive estimators. For off-line estimation, four fully constrained least-squares methods were compared. The results showed that the variations of a constrained least-squares approach produced more statistically efficient estimates. For on-line estimation, two recursive least-squares algorithms were examined. The first method extended Kalman filtering to satisfy the natural constraints of the O-D split parameters (the proportion of traffic from a particular on-ramp exiting at other on-ramps must sum to 1). The second was developed from sequential quadratic programming. These algorithms showed different capabilities to capture an abrupt change in the split parameters. The off-line method has great appeal due to its simplicity, Levinson says. Also, the method does not need an a priori start solution—multiple starting solutions are generated using the data set. The greater the ability of AIMSUN to match the actual model that generated the data set, the more performance improves. Experiments on the first test site yielded interesting findings related to mainline dominance and identifiability. Modifications to the traditional objective function have been proposed and the two-step optimization seems promising. Also, the use of a micro-simulation makes the evaluation of the estimates better because counts as well as other system statistics can be compared.
What's next?
Future work could enhance the performance of the microscopic traffic simulator with better calibration, Levinson says. In addition, the new modifications to the traditional objective function, the two-step optimization, and the new objective function need to be investigated. Finally, the identifiability issue of insufficient information in the off-ramp counts can be investigated by experimenting with alternate sites that have additional information, like a small network with turning-movement volumes (at intersections, traffic at a point may go three ways—forward, left, or right—as opposed to a freeway network where at most junctions drivers have only two choices—exit or do not exit).