Seminars / Conferences

Spatial mismatch, i.e., location misalignment between where supply and demand arise, is a common source of inefficiency in matching applications such as ride-hailing. In ride-hailing, this misalignment often leads drivers to travel long distances to pick up riders. First, we showcase how spatial mismatch can be mitigated through the commonly used operations concept of pooling (i.e., aggregating demand spatially or temporally) and, perhaps surprisingly, depooling (i.e., restricting matches to be more local, such as zoning and localization). Second, with temporal/spatial pooling, we study the optimal staffing problem for an on-demand vehicle-sharing firm operating over a d-dimensional service region. The firm makes a one-time capacity decision of how many vehicles to deploy and dynamically controls vehicle-customer matching and vehicle routing over time. The objective is to minimize the long-run average cost, which includes vehicle operations costs and customer waiting costs. We show that, jointly with optimal matching and routing, the optimal staffing level consists of a nominal load (i.e., the absolute minimum number of vehicles to ensure system stability) plus a safety buffer. This safety level depends on key system parameters, including the dimensionality d, the distributions of customer origins and destinations, and the customer-pooling capacity q (i.e., the maximum number of customers per vehicle). Specifically, (i) when q=1, the safety level scales with the customer arrival rate raised to the power of d/(d+1), a result that mirrors earlier findings under stylized spatial models. (ii) When q>1, the scaling becomes 2d/(2d+1), which, to our knowledge, was not discovered in the past.

The rapid integration of AI into educational contexts presents opportunities and risks. This talk reports findings from two field studies examining how students interact with AI and the subsequent impacts on learning. In the first study, we implemented a 12-week chess training program to investigate repeated exposure to AI-assisted learning. While student autonomy is often assumed to support agency, we find that giving learners full control over when to request AI assistance diminished engagement and halved performance gains. The immediate convenience of help displaced the productive struggle required for skill development, though these negative effects were fully mitigated among highly motivated students but not among highly skilled ones.

The second study, conducted with 800 first-year computer science students, tested a targeted intervention: training with “adversarial examples”—problems designed to elicit incorrect responses from ChatGPT. Students who received this training developed stronger skills in identifying and correcting AI errors, enabling more effective human–AI collaboration. Taken together, these studies suggest that simply providing students with unrestricted access to AI tools can be harmful. However, structured interventions that teach students to critically evaluate AI outputs hold promise. [Based on joint work with Stefanos Poulidis, Angel Chung, and Osbert Bastani.]

This research investigates the ability of large language models (LLMs) to generate and evaluate
new product ideas. Across a series of studies, we highlight both the strengths and limitations of LLMs in product innovation. In our first study, we show that LLM-generated ideas achieve higher average quality than human ideas (measured by purchase intent) and are seven times more likely to rank in the top 10%. A second study demonstrates that this creativity advantage is not simply due to the LLM’s superior persuasive skills. Our third and fourth studies reveal a key weakness: AI-supported brainstorming produces ideas that are less novel at the individual level and less diverse at the set level. Finally, in our fifth and sixth studies, we shift focus from idea generation to evaluation, exploring how LLMs can be used to forecast the success of new
opportunities.

Online rental platforms that allow customers to purchase rented goods present a complex pricing
challenge: setting buyout prices that balance immediate sales revenue with future rental income, while accounting for item-specific factors such as condition, popularity, and customer preferences. In this
paper, we develop, estimate, and validate a data-driven framework to inform buyout prices in this setting. Leveraging a Markov Decision Process (MDP), our framework assesses individual item value
based on rental demand, product attrition, and customer purchase likelihood. We use real-world data
from a leading fashion rental company to demonstrate that our methodology significantly improves profitability compared to existing practices and alternative benchmarks. We estimate that the proposed
pricing policy increases earnings by 3.1% over the company’s current practice. Our analysis also shows
that operating a rental-only business model leaves revenue opportunities untapped, underscoring the
strategic value of buyout options in managing inventory and generating additional income.

Queueing networks are foundational for analyzing service systems, but their construction has traditionally required substantial domain expertise and manual specification of arrivals, service rules, and routing logic. This talk presents a data-driven framework that leverages autoregressive sequence models to automate the development of queueing network simulators directly from event-stream data. Our approach learns the conditional distributions of event types and times, reframing the modeling task as one of sequence distribution learning. We demonstrate that Transformer-style architectures can effectively parameterize these distributions, enabling automated construction of high-fidelity simulators. As a proof of concept, we validate our framework on event tables generated from diverse queueing networks, highlighting its utility for simulation, uncertainty quantification, and counterfactual evaluation. By harnessing advances in machine learning, our framework lowers the barrier to applying queueing models and paves the way for scalable, data-driven decision support across a wide range of service operations.

Maritime transportation accounts for 90% of global trade, but ballasting—vessels traveling without cargo—imposes substantial economic and environmental costs. This paper examines the oil transportation industry, where approximately half of all miles traveled are sailed empty. While some ballasting is necessary due to inherent supply-demand imbalances in oil markets, our analysis demonstrates that market structure, specifically the fragmentation of vessel ownership, is also a primary driver, accounting for 10-20% of the total empty miles traveled depending on the market segment. In addition, we show that consolidating vessels into small shipping pools—sets of vessels operated under unified management—can reduce ballasting-related carbon emissions by up to 15%. This market-driven approach, which is gaining industry adoption, maintains competitive dynamics, given the limited scale of consolidation, while significantly improving efficiency. The gains arise from enhanced coordination within larger pools and expanded port coverage, reducing unnecessary vessel repositioning. More broadly, our findings quantitatively demonstrate that organizational changes alone—specifically, the consolidation of vessel operations—can generate significant environmental improvements by reducing empty miles. This provides a practical path toward sustainability that can complement and amplify the benefits of technological innovation.