These are the various academic research projects that I am playing around with. The level of interest varies among the topics.

Travel time estimates from single-loop detectors

As advanced traveler information systems become increasingly prevalent the importance of accurately estimating link travel times grows. Unfortunately, the predominant source of highway traffic information comes from single-loop loop detectors which do not directly measure vehicle speed. The conventional method of estimating speed, and hence travel time, from the single-loop data is to make a common vehicle length assumption and to use a resulting identity relating density, flow, and speed. Hall and Persaud (1989) and Pushkar, Hall, and Acha-Daza (1994) show that these speed estimates are flawed.

In this research we investigate methods of estimating link travel times directly from the single-loop loop detector flow and occupancy data without heavy reliance on the flawed speed calculations. Our methods arise naturally from an intuitive stochastic model of traffic flow. We demonstrate by example on data collected on I-880 data that when the loop detector data has a fine resolution (about one second), the single-loop based estimates of travel time can accurately track the true travel time through many degrees of congestion. Probe vehicle data and double-loop based travel time estimates corroborate the accuracy of our methods in our examples.

This work is done in collaboration with professors Peter Bickel, and John Rice, both from the Department of Statistics at UC Berkeley, as well as their graduate students Mike Ostland, and Xiaoyan Zhang.

Optimal Placement of FSP Tow Trucks

Freeway service patrols (FSP) are a popular means of incident management and control. In this research we address the question of the correct placement of FSP FSP tow trucks as a scarce resource allocation problem. We investigate methodologies for determining where to place FSP tow trucks so as to maximize the expected reduction in congestion. We illustrate this approach using the I-880 database. We note that any attempt to quantify the optimal placement strategy with regards to using real data will ultimately fail (but apparently will still be presented at TRB).

This work is being done with Dr. Alex Skabardonis and Prof. Pravin Varaiya.

Methodology for the Evaluation of FSP Tow Trucks

This research encompasses the Freeway Service Patrol (FSP) Evaluation Project proposed for the Los Angeles area. The goal of this project is to determine the cost effectiveness of the FSP program in Los Angeles where they have higher flows on the freeway. While the types of data collected during the proposed study will be roughly the same as the data collected for the Bay Area FSP Study, there will only be one set of data collected. We discuss the various assumptions that are made by using this different methodology. A short study is done with the data collected during the Bay Area FSP Project that shows that the benefit to cost ratio obtained by using this proposed methodology is only 0.53. This should be compared to the benefit to cost ratio of 3.08 that was calculated with the methodology used during the Bay Area FSP Study. Some suggestions for the LA FSP Study are given in the last section.

The main point of this research is that the LA FSP proposal only includes one study; one where the FSP tow trucks will already be in operation. Consequently, it is difficult to obtain accurate values for the assisted incident durations and their respective delays when there are no FSP tow trucks in operation. In this research we present and discuss a methodology that will allow us to estimate these quantities based only on ``after'' study data. Hence, this methodology presents a way to calculate the benefit to cost ratio for the LA FSP study. We test this methodology with the data from the Bay Area FSP Project.

This work is being done with Dr. Alex Skabardonis, Prof. Pravin Varaiya, and Robert Bertini.

The FSP Homepage contains more information on this project.

Incident Detection with Probe Vehicles

In this research we develop an incident detection algorithm based on information received in real-time from probe vehicles. We investigate models which allow us to estimate the upper bound detection rate for a given density of probe vehicles. We demonstrate our algorithm on data collected from the I-880 freeway in Hayward, California. We observe that a probe vehicle-based algorithm is feasible, and it avoids some of the infrastructure problems facing loop-based algorithms.

Sensor Fusion for Incident Detection

The cost of delay on freeways caused by non-recurring incidents is significant. Some estimate that the cost will be $35 billion/year by the year 2005 (Lindley, 1986). To reduce the impact of an incident a traffic management center (TMC) needs to quickly detect and remove it from the freeway. In this vein a large amount of research has been spent on the quick detection of incidents. Since in a large urban environment the automation of this task is crucial, automatic incident detection algorithms have been the subject of study now for more than 20 years.

Most research efforts in this area have dealt with trying to interpret information obtained solely from loop detectors. The developed algorithms have varying degrees of success with respect to detection rate, false alarm rate, and the mean time to detect an incident. Unfortunately, the common problem with all of the current algorithms is the high rate of false alarms that make them problematic to use in a large urban environment.

In this research we propose a novel algorithm for detecting incidents on the freeway that uses not only loop detector data but also data from mobile reporting sources. Mobile sources include FSP and CHP reports as well as cellular phone calls. These two diverse sources of information, each with their own degree of resolution and reliability, are combined using data association and sensor fusion techniques to enhance the decision-making capabilities of our incident detection algorithm. The incident detection algorithm that we have developed collects the data, processes it, reduces the uncertainty in it and then produces an inference about an incident.

With real data we demonstrate that by using these mobile reports to adjust the a priori probabilities of there being an incident we can achieve not only better false alarm rates but also better detection rates than conventional loop detector-based algorithms. In some sense this adjustment of the probabilities is simply making the algorithm more or less sensitive to detecting an incident based on the number of type of mobile reports (CHP reports are more reliable than cellular phone reports). We also investigate the complexity involved with translating the cellular phone calls into a form acceptable for a sensor fusion algorithm.

We conclude that a sensor fusion-based algorithm is practical and desirable. Indeed it is intuitive that one should take into account all of the information that is available to a TMC when trying to detect incidents.

Software Architecture for ATMIS Applications

There is a significant amount of static and dynamic data on travel conditions in typical highway network. Within the framework of ITS, there have been many studies on the types of services that should be provided to users. We examine the software design of a few ATMIS systems and identify limitations in their design and implementation. We propose that a centralized data collection and processing scheme will provide users with the most functionality for the least cost. It will also provide system designers with a controlled framework for improvements and upgrades. We propose ways for data structure, processing, fusion and presentation to both traffic managers and system users. This architecture is a possible starting point for the deployment of other National ITS Architecture systems. Finally, we investigate one possible implementation which we feel addresses the problems of data management and processing effectively.

Optimal Load Balancing in Hetrogeneous Networks with Sparse Information

Numerous studies have demonstrated the benefits of load balancing in a heterogeneous distributed local area network. These studies always start from the assumption that the user has complete control over the setup and administration of the machines in their cluster. This implies that programs can be installed on all the hosts, prior to the execution of the user tasks, that will assist in remote execution and/or migration of the jobs and, more importantly, will cooperate in idle resource detection. We feel that this assumption is too restrictive because very few researchers have access to an installed, cooperating process distribution system. The xdistribute process distribution system allows users to distribute jobs to remote hosts without any remote host software installation, administrative support or host cooperation. It knows nothing about the speed or load on any of the remote hosts until after the system has started distributing processes.

The problem that we address is how to optimally distribute tasks to remote hosts under these conditions. Under these assumptions the most restrictive factors become the time that it takes to setup the remote host to accept jobs, the time to install any custom software (we don't assume a uniform file system), and the time that it takes to transfer the task. We present an adaptive distribution scheme that works surprisingly well with no prior knowledge of the resources available and no cooperation for idle resource detection. We demonstrate conditions under which this adaptive scheme performs close to the schemes that use cooperation for idle resource detection. Experimental results show that this distribution schedule, working with the xdistribute distribution system, is an efficient and simple way to take advantage of idle hosts in a heterogeneous workstation environment.