UC Irvine

This research seeks to understand how information displayed by variable message signs (VMS) can affect driver route-choice and be better used for active traffic incident management. I study the effect of various VMS messaging strategies using a money incentivized behavioral experiment with a novel 2D real-time driving simulator that supports dozens of subjects driving on a shared virtual roadway where traffic incidents unpredictably occur. Drivers are shown a VMS display before choosing between two congestible routes. I conducted this experiment with students at the UCI Experimental Social Science Laboratory (ESSL) and with a more diverse sample of online subjects crowdsourced from the Amazon Mechanical Turk (MTurk) marketplace.

Chapter 1 will present the research motivation and methodology, the design and implementation of the experiment platform, and the results with student subjects. I find that subjects learned to efficiently operate the driving simulator, all tested VMS messaging strategies improved aggregate outcomes compared to the No VMS baseline, displaying messages didn’t cause highly volatile diversion rates, and subject gender exhibited consistent correlations with route choice.

Chapter 2 will discuss the reasons for replicating on MTurk, the methodological modifications necessary to conduct the experiment online, and how the MTurk results compare to the student results. I find that it’s viable but challenging to conduct real-time multiplayer experiments on MTurk, there are significant differences in individual characteristics between the MTurk and student subjects, and there are limited behavioral differences between the two groups.

Chapter 3 will introduce a framework using long short-term memory (LSTM) neural networks to predict driver route choice using real-time contextual data. I use varyingly limited vectors of data from my driving simulator experiments as the neural network’s input to predict driver route choice at the decision point between the two available routes. I find that the best performing model configuration can predict individual route choice with 74.0% average accuracy with in-sample cross validation and 72.2% average accuracy with out-of-sample validation.