Presentation details:

Title: Learning to Price with Limited Information

Date: May 5th, 2026

Time: 10:00 AM – 12:00 PM ET

Location: KACB 3100 (Zoom link)

Committee:

Prof. Sahil Singla (Advisor), School of Computer Science, Georgia Institute of Technology

Prof. Jan van den Brand, School of Computer Science, Georgia Institute of Technology

Prof. Santosh Vempala, School of Computer Science, Georgia Institute of Technology

Prof. Will Ma, Graduate School of Business and Data Science Institute, Columbia University

Prof. Rad Niazadeh, Booth School of Business, University of Chicago

Abstract:

Resource allocation is a central problem in theoretical computer science, operations research, and economics. In a resource allocation problem, a platform must design efficient algorithms to assign limited resources to participating agents under uncertainty, aiming to maximize either the total social welfare or the total revenue.

Among the various algorithms for resource allocation, pricing algorithms are one of the most common ways to distribute resources due to its simplicity. Theoretically, pricing algorithms can be modeled as a system in which a platform sets prices for resources, while participating agents with valuations drawn from underlying distributions make greedy decisions with respect to the prices. However, these distributions are typically unknown to the platform and must therefore be learned, either from limited offline data or through real-time interactions with agents.

Traditional approaches to learning in pricing algorithms often assume access to a large number of samples from each agent. In many real-world markets, however, the available information is sparse, noisy, or indirect. This raises a fundamental question: how can we learn effective pricing algorithms under such limited information constraints?

In this thesis, I address this question by studying the learnability of resource allocation algorithms under three highly restrictive information models:

1. Learning from a single sample, where the platform has access to only one sample from each agent’s distribution.

2. Learning from pricing queries, where the platform can only post prices and observe agents’ decisions as feedback.

3. Online learning with limited feedback, where the platform needs to adjust the algorithm dynamically with limited feedback.

I will demonstrate how pricing algorithms can be successfully learned under these frameworks. Together, these results provide a unified perspective on learning pricing algorithms under limited information, and show that strong performance guarantees are attainable even in highly information-constrained environments.