Learning Truthful Mechanisms without Discretization

TL;DR

TEDI introduces a discretization-free, neural network-based approach for learning truthful mechanisms, significantly improving efficiency and expressiveness in automated mechanism design without sacrificing performance.

Contribution

The paper presents TEDI, a novel discretization-free algorithm that guarantees truthfulness and full expressiveness using a new network architecture and training techniques, advancing automated mechanism design.

Findings

TEDI achieves competitive or superior performance compared to state-of-the-art methods.

The approach guarantees truthfulness, expressiveness, and dimension-insensitivity.

It is the first method to learn truthful mechanisms without outcome discretization.

Abstract

This paper introduces TEDI (Truthful, Expressive, and Dimension-Insensitive approach), a discretization-free algorithm to learn truthful and utility-maximizing mechanisms. Existing learning-based approaches often rely on discretization of outcome spaces to ensure truthfulness, which leads to inefficiency with increasing problem size. To address this limitation, we formalize the concept of pricing rules, defined as functions that map outcomes to prices. Based on this concept, we propose a novel menu mechanism, which can be equivalent to a truthful direct mechanism under specific conditions. The core idea of TEDI lies in its parameterization of pricing rules using Partial GroupMax Network, a new network architecture designed to universally approximate partial convex functions. To learn optimal pricing rules, we develop novel training techniques, including covariance trick and continuous sampling, to derive unbiased gradient estimators compatible with first-order optimization. Theoretical analysis establishes that TEDI guarantees truthfulness, full expressiveness, and dimension-insensitivity. Experimental evaluation in the studied auction setting demonstrates that TEDI achieves strong performance, competitive with or exceeding state-of-the-art methods. This work presents the first approaches to learn truthful mechanisms without outcome discretization, thereby enhancing algorithmic efficiency. The proposed concepts, network architecture, and learning techniques might offer potential value and provide new insights for automated mechanism design and differentiable economics.