Category Archives: Recent research

Recent research

Can LLMs help align transportation policy making?

Leave a comment

Transportation policy shapes how cities and regions grow, move, and thrive. Yet even the most carefully modeled policies can falter when they meet public opinion. Traditional approaches rely on mathematical optimization to identify “best” solutions under formal objectives and constraints—but these models often rest on rigid assumptions about behaviors and preferences. The result is a persistent gap between policies that look good on paper and those that communities actually support.

Our new study explores whether large language models (LLMs)—the same AI systems that power tools like ChatGPT—can help bridge that gap. LLMs are trained on vast troves of human language, giving them a rich contextual understanding of social values, trade-offs, and reasoning. Could such models serve as a new kind of decision-support tool, helping policymakers anticipate public preferences before a policy is implemented?

To test this idea, we built a multi-agent voting framework in which autonomous LLM “citizens” represent diverse communities within a large city. These agents participate in simulated referendums over transit policies involving three levers: sales taxes, transit fares, and driver fees such as congestion charges. Their choices are compared against the benchmarks of a standard travel demand model grounded in transportation economics.

The results are intriguing. Across cities and models, the LLM-based referendums produced collective preferences that broadly align with model-based predictions—except that the AI “voters” show a stronger aversion to taxes. GPT-4o agents tended to vote more consistently and decisively, while Claude-3.5 agents were more nuanced, yet both converged on similar priorities. Interestingly, their choices also shifted between cities: in Houston, for instance, GPT-4o agents favored lower taxes and higher driver fees than in Chicago.

These findings suggest that LLMs can mimic public reasoning in meaningful ways—offering policymakers a new tool to explore how communities might respond to complex, real-world trade-offs.

You may download a preprint here.

Recent research

How our means and values shape transit policy

Leave a comment

Few topics in urban mobility spark as much debate as fare-free transit (FFT). On one side, advocates argue that eliminating fares can guarantee the “freedom to move”, boost ridership, and make cities more sustainable. For example, Boston’s mayor Michelle Wu, a big fan of FFT, has implemented a limited FFT program in the
city. The idea also featured prominently in the platform of Zohran Mamdani, who has launched a closely watched bid
for the New York City mayoralty.    On the other, skeptics worry about overcrowding, misuse, and—above all—how to replace the revenue that keeps buses and trains running. Behind the heated arguments lies a deeper question: how do our moral values and financial means shape the “best” transit fare policy?

In our new study, How our values and means shape optimal transit fare policy, we set out to answer that question with a rigorous modeling framework. We designed a system that can evaluate not just the economics of different fare options—ranging from full-fare-free (FFF) to partial discounts to today’s standard fares—but also their equity implications. The model captures real-world features such as the “zero-price effect” (the psychological boost of free services), the cost savings from removing fare collection, and the operational dynamics of running a large transit system.

Applied to Chicago, our model led to some interesting findings. When money is plentiful, making transit completely free can indeed be both efficient and fair. But under tighter budgets, targeted discounts often do more to help low-income riders while keeping the system financially sustainable. Perhaps counterintuitively, we also found that giving planners unlimited resources doesn’t always lead to better outcomes—sometimes moderate constraints produce solutions that more people would support.

As cities everywhere grapple with post-pandemic ridership losses and strained finances, these findings matter. Fare policy isn’t just a technical choice—it reflects what we value as a society and how much we’re willing (and able) to pay to uphold those values.

Read the full preprint here.

Recent research

Bucket effect simplifies robust relocation

1 Comment

Relocating idle assets is a core operational challenge in the transportation industry, particularly for shared mobility services such as ride-hailing, car-sharing, and bike-sharing. The productivity and profitability of these services depend on effectively aligning the spatial and temporal distribution of available assets with fluctuating demand. This often requires routinely repositioning idle assets to anticipated high-demand areas—or “hot spots.” Although relocation is costly and non-revenue-generating in itself, it is essential for mitigating persistent supply–demand imbalances. Efficient relocation strategies are therefore critical to enhancing both user experience and operational performance.

The central challenge of relocation lies in the inherent uncertainty of demand. In ride-hailing and car-sharing markets, for example, passenger arrivals at specific locations vary not only by hour of the day but also from day to day. Such variability is only partially predictable, as travel patterns are often influenced by irregular events such as weather or accidents. To cope with this uncertainty, relocation strategies must be designed to be robust, protecting the service against extreme outcomes. Our study is motivated by the formulation and solution of such a robust relocation problem.

While the basic formulation of the problem is straightforward, solving it efficiently is not. We address this challenge with a novel insight inspired by the well-known “bucket effect.” Specifically, given a fixed relocation routing policy, once a zone is saturated with supply—meaning its demand is fully met at steady state—shifting additional demand elsewhere cannot increase overall profit. This insight, first conceived by Ruijie Li, the lead author and a long-term collaborator who visited my lab, provided the foundation for our approach.

You may read the abstract below or download the full paper.

Abstract: Relocating idle vehicles is a key challenge in autonomous shared mobility systems. This paper develops a robust optimization framework for steady-state relocation under uncertain demand, extending the fluid queuing network model of Braverman et al. [2019] to account for coupled uncertainty-where zone-level demand variations are bounded both locally and by a global budget. Although such uncertainty typically makes robust optimization intractable, we uncover a structural property, which we call the bucket effect: once a zone is saturated with supply (i.e., its demand is fully met) at steady state, shifting demand elsewhere cannot increase overall profit. This insight allows us to reformulate the robust relocation problem as a tractable linear program. Using New York City taxi data, we demonstrate the model’s practical value. The robust relocation strategy delivers higher profit and service availability than no-relocation baselines, while also providing greater stability under demand fluctuations. Importantly, these gains are achieved with sparse relocation patterns, indicating that targeted repositioning-guided by the bucket effect-can yield substantial benefits with modest operational effort.

Recent research

Pod for joint transit design

Leave a comment
Credit: NEXT Future

This paper explores a futuristic transit design concept enabled by modular autonomous vehicles (MAVs).   The goal is to explore the potential of jointly designing regular fixed-route transit service and paratransit.  The project was funded by CCAT, a US DOT Regional University Transportation Center headquartered at University of Michigan.  The leading author, Xiaoyu Yan, presented an earlier version of the paper at 2025 Transportation Research Board Annual meeting.   For preprint, please check SSRN.  The abstract follows.

Abstract  This study envisions a jointly designed transit system comprised of a fixed-route (FR) service and a paratransit (PT) service. The integration of the two services is inspired by the potential application of modular autonomous vehicles, or pods, in transit. Constrained by a fixed budget, the operator of the joint system aims to minimize the total user cost by optimally allocating pods between the two services. To formulate the operator’s design problem, we propose a stylized model, in which the FR service features a simple 2D grid route structure overlaying on a square city, and the PT service is designed as a general on-demand system that can be configured in different modes of operations. A case study is conducted using transit data from the Chicago region. We find that joint design helps prevent resource misallocation that could render a service dysfunctional under insufficient budgets, although its potential to reduce total user cost is limited. Enforcing the equal-access constraint—requiring that PT users incur no greater cost than FR users—tends to help PT users at the expense of FR users, though the overall impact on total user cost is insignificant. Modularity enables the formation of pod trains using small pods, which benefits FR operations, particularly when the design is not tightly constrained by budget. In contrast, automation delivers greater service improvements for PT users, whose more labor-intensive cost structure makes them more sensitive to efficiency gains, especially under tight budgets. Among the PT service modes, ridesharing is the most flexible, allowing for a wide range of service levels based on the available budget.

Recent research

Information design

Leave a comment

I have been collaborating with my PhD student, Qianni Wang, on information design for over a year. This paper—currently under review at Transportation Science—is the first fruit of our dedicated effort, and I am confident it will not be the last.

At its core, the idea behind information design is both simple and fascinating. If you are not yet familiar with it, I encourage you to explore Bayesian Persuasion, one of the earliest and most influential works published on the subject.

Information asymmetry—where one party holds information unavailable to others — lies at the heart of many important problems in economic and sociotechnical systems . In principal-agent problem (or the incentive problem), for example, the interests of the principal and agents diverge because the principal cannot perfectly monitor agents’ hidden actions or intentions. In the selection problem, one party in a transaction has private information that can lead to mismatches or undesirable outcomes. The beauty contest (or social learning) problem—which causes socially harmful herd behaviors—arise from individuals’ imperfect information, both of the system state and of what others know. The adverse effects of information asymmetry may be mitigated by aligning incentives, increasing transparency and inducing truthful behavior, often achieved through mechanism design.

Information design is motivated by information asymmetry that strongly favors the principal over the agents in a principal-agent game. In the context of transportation management, the principal is the manager and the agents are the users. Here, the users’ payoff depends on certain information observable only to the manager. As a result, the manager can exploit the information asymmetry to better align the interests of the two parties through persuasion . A unique feature of information design is that the manager is committed to an information structure ex-ante, which provides transparency and ensures incentive compatibility of the users.

This paper applies the concept of information design to a particular transportation application: resolving chaos in EV charging.  You can find an abstract below, and download a preprint here.

Charging remains an obstacle to the mass adoption of electric vehicles (EVs), especially for long-distance travel. If many EV drivers take to the road roughly at the same time, their “range anxiety” may create a self-fulfilling prophecy. As the drivers anticipate uncertainty and congestion at charging stations, they tend to make overly conservative decisions about when and where to charge. Yet, these decisions could worsen the very problem they try to prevent, collectively leading to chaos and inefficiency. Here, we show that information design can be used to persuade the drivers to adopt decisions that are better for the system while being consistent with their self-interest as defined by Bayesian Nash equilibrium. Our stylized model incorporates the congestion effect into the drivers‘ payoffs. It also assumes that the information designer has private knowledge about the state of charging and the driver type, defined by vehicle range. We consider both public and private information designs. The former does not depend on the driver type while the latter does. For the private design problem, we propose a novel cutoff structure that enables us to reformulate an infinite-dimensional problem as a finite one. When the random charging state can only take two discrete values, we prove that the optimal public design equals full information revelation. The optimal private design, however, promises significantly better results, potentially delivering the system optimal outcome under favorable conditions — e.g. when the charging state is highly uncertain and the marginal cost is similar regardless of the charging decision. We also show that the value of information increases with the level of uncertainty and the extra cost imposed by charging.

Recent research

What cellphone data reveal about teleworking

Leave a comment

My student Tianxing has been working hard to decipher cellphone data for some time now. Earlier this year, we have completed a paper, in collaboration with Amanda’s group,  showing unexpected representative biases in cellphone data that appears to have direct link with privacy regulations.  In this study, we uncovered work types of cellphone users using a clustering algorithm, and validated the results against  surveys data and regression analysis.  You may download a preprint here. The abstract follows.

In a short period, the COVID-19 pandemic has transformed telework into a common practice for a significant portion of the workforce. This shift has profound implications for land use, urban development, and transportation. Traditional survey-based methods for tracking these changes are struggling to keep pace with the rapidity of this transformation. Here, we propose a method to identify different types of workers from mobile phone data, which allows us to closely examine the correlation between work arrangements, mobility patterns and key socio-demographic attributes. By applying a hierarchical clustering algorithm to a set of features extracted from a mobile phone data set, six different work types are identified and their validity is confirmed using different approaches. We find teleworkers tend to travel slower than regular workers but faster than non-workers. They also travel a shorter distance to reach their primary activity location than regular workers, but a longer distance to reach other activity locations than both regular and non-workers. Our regression analysis further shows that, largely in agreement with findings in literature, racial minority and low income groups are less likely to telework. Implications for the use of trace data to model the evolving relationship between mobility and worker-classification are discussed.

Recent research

Is Fare Free Transit Just?

Leave a comment

I became interested in fare free transit since  Michelle Wu was elected the Major of Boston. She was the first female Asian major of the city, though her reputation as a disciple of Elizabeth Warren, the liberal firebrand in the U.S. senate,  probably overrode her other identities.  Among many of her agenda items was fare free transit (FFT), which caught my attention  not because it is especially progressive, but because it is a transit policy, which I happen to know something about.  Another source of inspiration for this paper came from Steven Dubner’s podcast on the subject a couple of years ago, which is entitled “Should Public Transit be Free”.

I shared the preprint with my department chair, Prof. Kim Gray, who is an environmental engineer but has a broad interest in anything related to sustainability and climate change. She was impressed and asked her assistant, Miss. Gina Twardosz, to write a news article to be posed on the department website. If you don’t want to read the paper itself, here is the link to that article. The abstract follows.

Abstract: Using a stylized transit design model, this study examines fare-free transit (FFT) through the lens of distributive justice. We pose a direct question: Is FFT just according to John Rawls’s theory of justice? Specifically, is it compatible with the resource allocation that maximizes the utility of the most disadvantaged travelers? We compare this egalitarian principle with a utilitarian one, which asserts that an allocation is optimal when it maximizes the total utility of all travelers. FFT is of course not free. In the absence of farebox revenue, a transit system must either cut services or turn to alternative sources, such as local dedicated taxes and fees levied on drivers. Thus, our model incorporates both finance and operational decisions, and captures the interaction between traffic congestion and travelers’ income level and mode choice. Using a case study built with empirical data in Chicago, we show that fare is not the first choice under either moral principle. For the egalitarian, the most desirable funding source is the driver fee, whereas taxation is preferred by the utilitarian. It follows that FFT can be both just and utility-maximizing, if one is allowed to raise taxes and charge drivers with impunity. However, as the flexibility in finance diminishes, so does the appeal of FFT. In such cases, the proposed model serves as a decision-support tool for finding sensible compromises that address the varied interests and ideologies at play. For example, it reveals that at the current tax rate of about 1% in Chicago, the Rawlsian egalitarian can justify FFT only if drivers pay about $1,800/year to fund transit, which amounts to about 18% of an average U.S. household’s driving cost.

Recent research

Unexpected Data Bias in Smartphone Trace Data

2 Comments

This study, a joint study with Professor Amanda Stathopoulos‘ group, explores the impact of shifting device representation bias in smartphone tracking data collected before and after Apple’s 2021 privacy updates on user location tracking. It demonstrates that privacy regulations can significantly and unexpectedly affect the quality of these data, which are crucial for decision making across governmental, corporate, and academic institutions worldwide. The research also corrects misconceptions about representation bias previously speculated in the literature. Overall, the findings equip users of location-based device data with a better understanding of potential pitfalls, enabling them to anticipate the changes caused by the evolving regulatory landscape and to devise appropriate coping strategies. This finding is contrary to popular concerns about the under-representation of low-income populations in LBS data.

Download the preprint here and read the abstract below:

As smartphones become ubiquitous, practitioners look to the data generated by location-tracking services enabled on these devices as a comprehensive, yet low-cost means of studying people’s daily activities. It is now widely accepted that smartphone data traces can serve as a powerful analytical tool for research and policymaking. As the use of these data grows, though, so too do concerns regarding the privacy regulations surrounding location tracking of private citizens. Here, we examine how Apple’s tightened privacy measures, designed to restrict location-tracking on their devices, affect the quality of passively generated trace data. Using a large sample of such data collected in the Chicago metro area, we discover a significant drop in iOS data availability post-privacy regulations. The results also reveal a surprising puzzle: the reduced tracking is not uniform and contradicts customary concerns about the under-representation bias of low-income population. Instead, we find a negative correlation between device representation level and income, as well as population density. These findings reframe the debate over the increasing reliance on smartphone data, highlighting the need to understand evolving issues in tracking, coverage, and representation, which are essential for the validity of research and planning.

Recent research

Entropy maximization for multi-class assignment

Leave a comment

The lack of uniqueness constitutes a serious concern for any analysis that relies on class-specific traffic assignment results, such as understanding the impact of a transport policy on the welfare of travelers from different income groups, sometimes known the vertical equity analysis.  Entropy maximization is a standard approach to consistently selecting a unique class-specific solution for multi-class traffic assignment.

Here, we show the conventional maximum entropy formulation fails to strictly observe the multi-class bi-criteria user equilibrium condition, because a class-specific solution matching the total equilibrium link flow may violate the equilibrium condition. We propose to fix the problem by requiring the class-specific solution, in addition to matching the total equilibrium link flow, also match the objective function value at the equilibrium.  This leads to a new formulation that is solved using an exact algorithm based on dualizing the hard, equilibrium-related constraints.

Our numerical experiments highlight the superior stability of the maximum entropy solution, in that it is affected by a perturbation in inputs much less than an untreated benchmark multi-class assignment solution.  In addition to instability, the benchmark solution also exhibits varying degrees of arbitrariness, potentially rendering it unsuitable for assessing distributional effects across different groups, a capability crucial in applications concerning vertical equity and environmental justice. The proposed formulation and algorithm offer a practical remedy for these shortcomings.

This is the third paper completed by the first author, Qianni Wang, who officially joined my group last year.

The paper was currently under review at Transportation Research Part B.  You may download a preprint here, or read the abstract below.

Abstract: Entropy maximization is a standard approach to consistently selecting a unique class-specific solution for multi-class traffic assignment. Here, we show the conventional maximum entropy formulation fails to strictly observe the multi-class bi-criteria user equilibrium condition, because a class-specific solution matching the total equilibrium link flow may violate the equilibrium condition. We propose to fix the problem by requiring the class-specific solution, in addition to matching the total equilibrium link flow, also match the objective function value at the equilibrium. This leads to a new formulation that is solved using an exact algorithm based on dualizing the hard, equilibrium-related constraints. Our numerical experiments highlight the superior stability of the maximum entropy solution, in that it is affected by a perturbation in inputs much less than an untreated benchmark multi-class assignment solution. In addition to instability, the benchmark solution also exhibits varying degrees of arbitrariness, potentially rendering it unsuitable for assessing distributional effects across different groups, a capability crucial in applications concerning vertical equity and environmental justice. The proposed formulation and algorithm offer a practical remedy for these shortcomings.

Recent research

Is competition for losers in bikesharing?

Leave a comment

The rise and fall of the bikesharing industry in China offers a cautionary tale about the risks of an unregulated market with a low entry barrier. It is well known that, while low entry barriers can promote competition and innovation, they may also lead to higher market volatility and potential challenges in achieving profitability due to intensified rivalry . There are also limited economies of scale to be had, making it exceedingly difficult to establish a monopoly. As Peter Thiel noted, “competition is for losers”‘ in such markets and good entrepreneurs should simply stay away from them.   However, writing off the bikesharing industry as unprofitable cannot be the only story here. After all, bikesharing has a genuinely positive societal impact and should have its place in many of our cities that are haunted by the disease of auto-dependency. The question is what, if anything, can be done to foster a healthy bikesharing market that is attractive to both users and private investors.  We set up to answer this question here.  You may download a preprint here, or read the abstract below.

Abstract: We model inter-operator competition in a dockless bikesharing (DLB) market as a non-cooperative game. To play the game, a DLB operator sets a strategic target (e.g., maximizing profit or maximizing ridership) and makes tactical decisions (e.g., pricing and fleet sizing). As each operator’s payoff and decision set are influenced by its own decisions as well as those of its competitors, the outcome of the game is a generalized Nash equilibrium (GNE). To analyze how competition may shape the choice of strategic targets, we further augment the game framework with a ranking scheme to properly evaluate the preference for different targets. Using a model calibrated with empirical data, we show that, if an operator is committed to maximizing its market share with a budget constraint, all other operators must respond in kind. Otherwise, they would be driven out of the market. When all operators compete for market dominance, Moreover, even if all operators agree to focus on making money rather than ruinously seeking dominance, profitability still plunges quickly with the number of players. Taken together, the results explain why the unregulated DLB market is often oversupplied and prone to collapse under competition. We also show this market failure may be prevented by a fleet cap policy, which sets an upper limit on each operator’s fleet size.