The following is my (Fisher) description of Dr. Rajeev Saraf's dissertation. I coadvised Rajeev with Professor Mark Abkowitz of Vanderbilt's Department of Civil and Environmental Engineering. The dissertation is a finalist for the Milton Pikarsky Memorial Award for best PhD dissertation in Transportation for 1994.


Rajeev's dissertation evaluates artificial neural networks for responsive, adaptive signal-plan generation. Responsive plan generation is a much discussed, but little implemented idea. The dissertation's basic premise is that existing signal-plan generation tools (e.g., TRANSYT) make rational decisions about signal plans under varying conditions, but these tools cannot be used in a real-time setting. An artificial neural network trained to approximate the decision-making behavior of such tools, however, can be used for real-time control that adapts to varying traffic demands. Artificial neural networks are computational devices that implement a form of multivariate regression or function approximation. They are currently simulated by computer software. Once `trained', response to environmental inputs is very fast; research efforts worldwide are implementing these types of devices in computer hardware, thus further decreasing computation time.

The dissertation frames existing work in signal control along two dimensions: the computer control architecture (i.e., centralized, hierarchically distributed, radically distributed) and the signal-plan selection strategy (i.e., pre-timed, selection from an existing plan library, online plan generation). The dissertation's niche is the evaluation of neural networks for online plan generation for centralized and radically-distributed (i.e., one controller for each intersection) control. The dissertation recognizes that these are architectural extremes, but that simplifying aspects of each architecture also make them the best starting points in a systematic exploration of the merits of a new traffic control technology.

For each architecture, Rajeev secured actual traffic flow data from appropriate sources: e.g., traffic flows at accumulated 15min intervals over 24 hours (96 data points) from a nine-intersection arterial to compare centralized plan-generation strategies. The amount of data was modest, and claims are conservative.

For centralized control comparisons the dissertation implicitly assumes that TRANSYT is used to determine signal plans at 15min intervals. The recommendations of TRANSYT are also used to train an artificial neural network. Ideally, training results in a neural network that responds in a manner similar to the trainer (i.e., TRANSYT) under similar circumstances. Experimental findings were that the trained neural network performed comparably to TRANSYT in terms of standard performance indices on traffic patterns that were not used to train the network. To simplify somewhat, the trained neural network reasonably generalized the output behavior of TRANSYT. In fact, the neural network performed slightly better than TRANSYT in terms of performance (though not significantly so). Rajeev suggests some interesting reasons why the neural network learner shows any performance improvement at all over its trainer. The results also illustrate some emergent properties of neural network control: it naturally `smooths' signal plan parameters such as cycle length, thus decreasing the difference between consecutive time points when compared to the signal plan recommendations that would be made by TRANSYT at 15min intervals. This has important implications for transition cost. Finally, the dissertation experiments with many variations on the basic neural network architecture, suggesting some interesting differences between various architectural choices, but in general, suggesting that the approach is `successful' across a wide range of neural network designs.

To assess the merits of a neural network controller in a radically-distributed environment, the dissertation uses OPAC as both a trainer and as a controller against which to make comparisons. In this setting both the neural network and OPAC are making decisions about when to terminate the current phase. The main finding is that the network outperforms OPAC along some important measures of effectiveness as measured under a NETSIM simulation. This again invites some discussion on how a `learner' can outperform its trainer. As before, Rajeev experiments with various neural network designs to qualify and more thoroughly test the basic approach. This includes an extension that trains the network to both terminate the current phase and select the next phase, with good results.