Neural network technology being applied to congestion management
Model may aid in predicting effects of TDM strategies
Congestion is slowing traffic and increasing pollution from Miami to Pensacola, and, if a CUTR research project succeeds, public agencies will be better able to predict how many vehicle trips—and resulting pollution—would be reduced by transportation demand management (TDM).
Through funding from the Florida DOT’s Research Ideas program, the application of neural network technologies to TDM is being researched by CUTR and the University of South Florida’s Department of Computer Science and Engineering. It is anticipated that the result will be the development of a neural network application for Florida to help streamline the development of trip reduction plans for employers and provide a basis for consistent review by the regulating agencies.
Neural networks are a group of highly-interconnected and relatively simple computational units that perform processing of inputs to produce a single output. The neural network connects the output of each unit to the inputs of many other units through different “weights.” Learning by a neural network consists of finding the correct number of computational units in the network with the correct numerical values of the weights that connect these units.
Anticipated results
The neural network model, which uses several years of data from the Los Angeles area, should help Florida with congestion management, growth management, and transportation demand management (TDM) efforts. The research team hypothesized that this application could improve the cost-effectiveness of the TDM programs and reduce administrative costs by:
- reducing the costs for developing and implementing plans to reduce vehicle trips by streamlining the plan development and review process for the regulated community,
- improving efficiency by reducing regulatory staff time in the review of employer/developer trip reduction plans,
- enhancing accuracy by developing consistent interpretations of plan submittals among these analysts, and
- making TDM programs more cost-effective.
Building the model
To conduct the model-building process, data from South Coast Air Quality Management District (SCAQMD) in California were compiled and “cleaned.” These data consist of more than 35,000 Average Vehicle Ridership (AVR) improvement plans submitted by companies in the greater Los Angeles metropolitan area. The research team then identified model inputs and outputs from SCAQMD data.
Travel impedance data that may influence the effectiveness of the plans were identified, collected, and entered. Although SCAQMD did not collect the data from employers, it was hypothesized that the data would be an important factor because travel distances have been shown to be significantly related to mode choices. This hypothesis was borne out when the travel impedance data were incorporated as an important part of the predictive models developed. The research team then designed and trained neural network models. The type of network that is most popular today is a multi-layer, fully-connected, feed forward neural network. This type of network consist of two or more layers of individual neurons, each of which, in a given layer, receives inputs from all the neurons in the previous layer. Its output is input to all neurons in the succeeding layer. The middle layers, between the input and output layers, are called “hidden” layers since they are not directly accessible. The sigmoid or the hyperbolic tangent allows each computational unit to implement a nonlinear mapping between its inputs and output. This allows the networks to carry out nonlinear relationships that may exist in the data.
Once the network is trained, it will produce the desired m-dimensional output given an n-dimensional input. The records provide the desired input-output associations used to train the network.
Results
The neural network that was developed has been found to be slightly more efficient in its use of input variables than alternative linear procedures (including standard multiple regression analysis and discriminant analysis). It also was able to correctly predict the impacts of substantially more plans than the Federal Highway Administration’s TDM Model, which is the only product currently available that attempts to perform the same function. It should be noted, however, that the FHWA TDM Model was originally designed to predict regional impacts (rather than impacts at individual sites) and that a substantial amount of data conversion had to be performed to produce FHWA model inputs from the SCAQMD data. Other data bases available from trip reduction programs that exist in Arizona, in particular, have been shown to have data very similar in design structure to the SCAQMD data. The next step in this project is to validate the model developed with the SCAQMD data by using it with data obtained from these sites in Arizona.
According to Ike Ubaka, Manager, Transit Systems Planning, Florida DOT, “We know that TDM strategies such as carpooling, vanpooling, and compressed work week programs hold the potential of reducing congestion and energy consumption. We are hoping to build a model to provide a tool for predicting the effects of combinations of these strategies.” For more information on the neural network project, contact CUTR TDM Program Manager Philip Winters at winters@cutr.usf.edu or Research Associate Francis Cleland at cleland@cutr.usf.edu, or call (813) 974-3120.
Please direct all comments and/or questions regarding this page to webmaster@cutr.usf.edu.
Return to the CUTR Home Page