My research focus is on how to build smart, seamless, efficient urban infrastructure to better serve society. I am interested in how we can measure existing systems, design new control systems and, importantly, swiftly transition to new systems. Specifically, I have focused my PhD work on studying new possibilities and limitations of machine learning, optimization, and control theory for improving transportation systems. My work is additionally guided by human-system interaction, public policy, and engineering ethics — as these are all crucial aspects in designing systems that aim to affect people positively one day.
Stabilizing highway traffic (Sep 2016–Current)
System-level energy reduction using team of autonomous vehicles via control theory and reinforcement learning
Researchers have been and continue to study the impacts of 100% human-driven vehicles or 100% robot cars on energy consumption. However, very little is known about what energy consumption looks like during the 35+ year transition period we face between 0 and 100% adoption of autonomous vehicles. This project aims to not only provide tools for studying such a transition but also propose controllers for enabling an ideal transition.
Scalable ridesharing (May 2015–Current)
Sustainable, adaptable, scalable mobility via clustering and local search
Carpooling has been long deemed a promising approach to better utilizing existing transportation infrastructure. This project studies optimization and behavior for carpooling systems, with the goal of designing a system that 1) is highly scalable, efficient, and adaptive, 2) people want to use, and 3) significantly decreases carbon footprint.
There are several reasons carpooling is still not the preferred mode of commute in the United States: first, complex human factors, including trust, compatibility, and not having right incentive structures, discourage the sharing of rides; second, algorithmic and technical barriers inhibit the development of online services for matching riders. High-occupancy vehicle (HOV) lanes which permit vehicles that hold three or more people (HOV3+) have been seen to simultaneously decrease trust concerns and dramatically reduce travel times, thereby providing a promising avenue for addressing both types of issues. One avenue of this project studies the complexity of the HOV3 Carpool problem and demonstrates that a relaxation of the problem is amenable to a wide range of common exact and heuristic methods for solving the problem of finding globally optimal carpool groups that may utilize these HOV lanes. Our experiments highlight that a hill climbing local search method scales up to 100K agents, thereby improving upon related previous work (which studies up to 1000 agents), and numerically converge to 1.1 of the lower bound.
In a related vein, by exploring alternative approaches to combinatorial optimization, we separately propose the first known formal connection between clustering and set partitioning, with the goal of identifying a subclass of set partitioning problems that can be solved efficiently and with optimality guarantees through a clustering approach. We prove the equivalence between classical centroid clustering problems and a special case of set partitioning called metric k-set partitioning. We discuss the implications for k-means and regularized geometric k-medians, and we present a case study in combinatorial optimization for ridesharing called the Rideshare Meetup Problem, in which we use an efficient Expectation Maximization (EM) style algorithm to achieve a 69% reduction in total vehicle distance, as compared with no ridesharing.
Traffic flow estimation (Sep 2013–Current)
Data-driven route flow estimation via block-simplex constrained signal recovery
Since we may only seek to control what we may first measure, one of my projects studies how well we can estimate traffic using technology available today. In this project we study the use of currently available cell tower data to estimate fine-grained flow information in traffic networks, which is difficult to measure directly. (Counter to common knowledge, GPS traces are a poor choice for estimating flow, though they are useful for estimating travel times.) Our convex optimization method scales well to full-sized networks, and our method enables better estimates for re-routing in the event of a road closure or accident, as well as better traffic light control. It also provides a tool for understanding how people actually use the road network; this serves as an alternative to utility and equilibrium-based models of road networks, which are widely studied but primarily designed for long-term land use planning and not short time horizon applications.
Given data from the cellular network, loop counts, license plate readers, and similar types of sensors on road networks, we estimate the link flow, origin-destination (OD) flow, and route flow of a road network. First, the OD demands are estimated by aggregating phone records by area and timestamp. Then, the problem of estimating route flows given the estimated OD demands and the sensor measurements is formulated as a quadratic program with block-simplex constraints. A change of variable reduces the dimensionality of the problem in which the simplex constraints become order constraints. We apply first and second order projected descent methods in which the projection is the solution of an isotonic regression problem with box constraints. Finally, we show that our method can be used to reconstruct link flows. We show that our algorithm scales well to the full-scale network of the greater Los Angeles area. We additionally benchmark our method against several other natural approaches for the block-simplex signal recovery problem, including compressed sensing and Bayesian inference. This project is joint work with Jérôme Thai, Steve Yadlowsky, Sara-Fleur Sultan, Prof. Alexei Pozdnoukhov, and Prof. Alex Bayen.
Additional areas of interest
- Online taxi pre-dispatch
- Online learning for resource allocation on networks
- Probabilistic guarantees on cyber-physical networks
- Two-sided market models for long-term city-scale demand modeling
- Distributed optimization for rebalancing of vehicle sharing systems
- Impact prediction and robust resource allocation for smart city disaster resilience