National
Institute of Statistical Sciences
19 T. W. Alexander
Drive
P.O. Box 14006
Research Triangle Park, NC 27709-4006
Tel: 919.685.9300
FAX: 919.685.9310
admin@niss.org
News: January, 1997
This five-year, $6 million project is funded by the Mathematics/Physical Sciences and Engineering Directorates at the NSF. It has created path-breaking collaborations between statisticians and transportation scientists and engineers, and is producing important results.
Currently, there are three primary research thrusts, all focused on road-based surface transportation: travel demand forecasting, intelligent transportation systems (ITS) and deterioration and durability of concrete. Alan F. Karr of NISS is Project Director; co-Principal Investigators are Eric I. Pas (Duke University) and Jerome Sacks (NISS). The project has involved more than 30 principal investigators, senior investigators and consultants; 6 postdoctoral fellows; and 10 research assistants from some 25 universities, research institutes, government agencies, government laboratories and corporations from the U.S. and abroad.
A principal concern of ITS is to employ modern sensor and information technology in order to use existing streets and highways more efficiently, by reducing congestion and travel times. Near-term examples include automated toll collection, variable message signs and Advanced Traveler Information Systems (ATIS), which furnish dynamic route guidance information to drivers. (In the longer run, fully automated highways are contemplated.) The markets are not only consumers, but also the commercial sector (taxis, trucks and package delivery services). However, much of ITS remains speculative, leading to critical issues of "what is needed" and "what works". Three subprojects have potentially very important implications for the assessment of ITS innovations. At the same time they illustrate the power and need for intelligent statistical analyses. We will describe these in some detail here. In later NEWS items we will discuss the work on travel demand and concrete.
Dynamic prediction of link and route travel times is essential to ATIS. Primary data sources are probe vehicles, which provide real time information on link travel times to a central controller; traffic signals, whose state is crucial to both values and dependence of link travel times; and single-loop detectors, which provide count (vehicles per time period) and occupancy (percentage of the time period during which there is a vehicle over the detector) at 5-15 minute aggregations. (Double-loop detectors also provide direct measurements of vehicle speeds, but they are rare.) Video data are increasingly common, but to date have been used only informally.
Fundamental questions, then, are what information is needed to predict travel times and what are the relative contributions of the different data sources. Specific issues, for example, the usefulness of probe vehicles as information sources (which may be infeasible on large scales), are of import as well.
To address these questions, NISS designed and conducted a field experiment on a small network in suburban Chicago during the summer of 1995. Data sources were probe vehicles, reporting link-by-link travel times (and also congested time and congested distance --- the time and distance traveled at speeds below a threshold), single loop system detectors and NISS-collected video tapes of the intersections at two ends of a "study link." In addition to signal status, the latter show exit times of all vehicles (not only probes). Key participants in conducting the experiment and analyzing the resultant data are Daryl Daley (Australian National), Alan Karr (NISS), Ashish Sen (Illinois at Chicago), Sim Soot (Illinois at Chicago), Todd Graves (NISS postdoc), Vonu Thakuriah (NISS postdoc) and students at the University of Illinois at Chicago and the University of North Carolina at Chapel Hill.
The most important conclusion from the experiment is that traffic signals and vehicle counts and occupancy are the key information needed to predict link travel times. The modeling process leading to this conclusion begins with simple models (for light traffic), which represent travel time as a function of entry time to the link (relative to the signal cycle). Then superimposed on these are traffic volume-dependent movement and queueing delays (including cycle failures). When signal status is available (from the signal system or video, for example) and when single-loop detectors are present to provide count and occupancy, there is little need for additional data (for example, from probes).
In reality, however, the "status-within-cycle" of signals is rarely available, whether from video or other sources. Therefore, techniques have been developed to impute signal status from travel time and congested time measurements provided by probes. The latter identify an unforeseen and potentially important role for probe vehicles.
An example of novel collaboration and new results comes from statisticians and transportation engineers at the University of California, Berkeley. Peter Bickel (California, Berkeley), John Rice (California, Berkeley) and Jaacov Ritov (Hebrew University, Jerusalem) and students from statistics and civil engineering at the University of California, Berkeley, have developed methods for estimating the distribution of travel times on a freeway using data from several single loop detector, using a very rich set of data collected by the PATH Project (Berkeley, CA). The PATH data set contains count and occupancy, at a time resolution of one second, separately for each lane, for some 20 locations on Interstate Highway 880 in Oakland, CA. Successive locations are approximately one-third mile apart. A critical step is dynamically to use the "naive" speed estimate (count divided by the product of occupancy and a nominal vehicle length) to center the estimated travel time distributions.
An important issue for management of freeways is recognizing incipient congestion. Karr, Nagui Rouphail (North Carolina State) and Graves, together with students from North Carolina State University, have developed methods to predict the onset of "freeway breakdown." This readily observed but not easily characterized phenomenon occurs when, in heavy traffic, freeways enter a regime of instability: without an external cause (such as an accident or lane closure), flows and speeds drop rapidly and dramatically, across all lanes. (Video data collected on I-40 in Research Triangle Park, NC, show speed drops from 60+ to 30 mph in less than two minutes.) From the standpoint of congestion management, the central need is to recognize breakdown before it occurs (in order to implement control measures such as changeable speed limits or access restrictions).
Over the summer of 1996, a unique data set was assembled by NISS (with cooperation of transportation engineers from the Washington State Department of Transportation): in real time and over the Internet, counts and occupancies were obtained from more than 30 single loop detector stations on a 10-mile stretch of I-5 in northern and north suburban Seattle, WA. Data were collected for weekday morning and afternoon peak periods, amounting to ten hours per day, over two months.
Visualizations of the data revealed a clear structure for breakdowns: a breakdown propagates upstream (as it is fed by demand) and clears downstream (as demand slackens). Thus, if breakdown as defined as "speed less than some threshold (e.g., 30 mph)", in space-time it has a characteristic triangular structure. Using this definition (with speed estimated "naively" from count and occupancy), CART-based methods have been constructed that predict breakdown with (from the perspective of traffic engineering) startling accuracy: using 10-minute blocks of data, more than 60 percent of breakdowns are predicted, with a false alarm rate of less than 5%.
Currently, in collaboration with Ronald Hughes of the Highway Safety Research Center of the University of North Carolina at Chapel Hill, the data are being used to attempt a similar analysis in order to identify conditions associated with accidents.
For details, see the Transportation Project Description.
Participants in the program will divide their time between NISS and one of the Departments or an affiliated Center. Participation at NISS will involve working with ongoing or prospective NISS research projects, for example, in transportation, software development, drug design, environment, education and large data sets. Departmental participation can range from teaching to involvement in ongoing research activities in, for example, Bayesian methods in imaging, neuroscience, and health policy; stochastic differential equations in economics and finance; extreme stresses on structures; biomathematics and atmospheric science.
A biweekly seminar series at NISS will draw together the multiplicity of activities and offer views of Triangle-based research, as well as interests related to NISS projects in other parts of the nation.
For details, see the Program Announcement, or send e-mail to admin@niss.org.
John C. Bailar, III, of the University of Chicago, was elected Chair of the Board of Trustees for 1997.
Dedication of the building will coincide with the 1997 NISS Annual Meeting, scheduled for November 7-8, 1997.
 |
 |
| Help |