Algorithmic Game Theory (AGT) studies systems where self-interested participants interact through algorithms and platforms, including ad auctions, online marketplaces, matching markets, and networked allocation. Classic AGT has been highly successful, but it relies on strong assumptions: independent private values, fully known priors, fully rational agents, clean structural properties (such as submodularity or gross substitutes), and worst-case analysis as the main benchmark. As articulated in the proposal, these assumptions often fail in real digital markets, making elegant theoretical guarantees fragile when deployed at scale. The RAGT project aims to systematically weaken these assumptions while preserving meaningful guarantees, in the spirit of Wilson’s doctrine.
The project addresses robustness along five axes:
(1) informational and outcome externalities, 2) unknown or misspecified priors, (3) behavioral considerations, (4) divergence from precise structural conditions, and (5) beyond worst-case inputs.
Across these axes, the goal is to design mechanisms that remain simple and detail-free when possible, and to quantify the performance cost when additional structure is required.
This agenda matters for society because modern economic activity is mediated by algorithms. Platforms for advertising, e-commerce, cloud markets, gig work, and digital services operate with correlated information, externalities, incomplete data, and users who deviate from idealized rationality. Mechanisms whose guarantees rely on fragile assumptions can misallocate resources, generate unfair outcomes, or become unstable in practice. By moving from “beautiful but brittle” models toward frameworks that incorporate correlation, imperfect priors, behavioral biases, and realistic input distributions, the project contributes tools for designing markets that are efficient, fair, and resilient to model error.
RAGT’s objectives are to identify which classical assumptions are truly necessary, design mechanisms that retain approximate efficiency or revenue when assumptions are relaxed, quantify degradation as models deviate from ideal benchmarks, and provide guidance on when simple robust mechanisms suffice and when more structured approaches are required. The project’s work on interdependent values, online allocation and prophet inequalities, fair division, noisy valuation classes, and learning-augmented settings aligns closely with these robustness dimensions.
