Learning-based Mechanism Design: Scalable, Truthful, and Continuum Approaches for Utility Maximization

Yunxuan Ma, Siqiang Wang, Zhijian Duan, Xiaotie Deng

Keywords: automated mechanism design, differential economics, function approximation, mechanism representation

TL;DR: We propose a neural-network based automated mechanism design framework that is not only intrinsically truthful, but also behaves well when the problem size is large.

Abstract:

Mechanism design is a crucial topic at the intersection of computer science and economics.
This paper addresses the automated mechanism design problem by leveraging machine learning and neural networks.
The objective is to design a **truthful**, **expressive** and **efficient** mechanism that maximizes the platform's expected utility, given that the players' types are drawn from a pre-specified distribution.

We present a general mechanism design model that captures two critical features: hidden information and strategic behavior.
Subsequently, we propose the **PFM-Net** framework, which parameterizes the menu mechanism class by function approximation and identifies an optimal mechanism through ingenious optimization techniques.
We also provide both theoretical and empirical justifications for the advantages of our approach.
Experimental results demonstrate the effectiveness of PFM-Net over traditional and learning-based baselines,
enabling the PFM-Net framework to serve as a new paradigm for automated mechanism design.

Primary Area: other topics in machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4223