UC Berkeley

Transportation networks such as road networks are well-known for their complexity. Its users make choices of route, which mode to take, etc.; these users then interact with each other, producing emergent dynamics such as traffic jams on roads. These localized multi-user emergent physical phenomena then interact with similar group movements occurring in other locations, creating more complex network-scale dynamics. These patterns of hierarchical levels of organization and emergent phenomena at each level are typical of so-called "complex systems." In addition, the increasing adoption of information-technology systems like connected and autonomous vehicles is creating new challenges in modeling transportation networks, as new emergent behaviors become possible, but also provide new sources of information and possibilities for traffic operations management.

The complexity of transportation networks precludes the use of a single all-encompassing theory for all situations at all scales. This dissertation describes several analyses into understanding and controlling emergent dynamics on road traffic networks. It is broken into three parts. The first part proposes models for several new phenomena at the "macroscopic," group-of-vehicles to group-of-vehicles, level. In particular, we solve a problem of modeling arbitrary road junctions with populations of behaviorally-heterogenous vehicles, where the vehicle flows are modelled by a continuum-approximation, partial-differential-equation-based model. We also present several new modeling constructions for a particular complex road network topology: freeways with managed lanes. It has been noted that these managed lane-freeway networks induce new emergent behaviors that are not present in traditional freeways; we propose modeling techniques for several of them, and fit them into traditional modeling paradigms.

The second part presents several contributions for estimating the state of the macro-scale traffic dynamics on the road network, based on the micro-scale data of global navigational satellite system readings of the speed and position of individual vehicles. These contributions are extensions of the particle filtering mathematical framework. First, we demonstrate the use of a Rao-Blackwellized particle filter in assimilating vehicle-local speed measurements to better estimate the macroscopic density state of a freeway. Then, we propose new "hypothesis-testing" particle filters that can be used to reject outlier or otherwise malign measurements in a principled statistical manner.

The third and final part presents two items on applying deep neural networks to transportation system problems at smaller scales. Both items make use of neural attention, which is a neural network design technique that allows for the integration of structural domain knowledge. First, we demonstrate the applicability of this technique towards estimating aggregate traffic states at the lane level, and present evidence that designing the neural network architecture to encode different types of lane-to-lane relationships (e.g., upstream lane vs neighboring lane) greatly benefits statistical learning. Then, we apply similar methods to an autonomous vehicle coordination problem in a deep reinforcement learning framework, and show that an attention-based neural network that allows each vehicle to attend to the other vehicles enables superior learning compared to a naive, non-attention-based architecture, and also allows principled generalization between varying numbers of vehicles.