Optimal auctions through deep learning

Abstract

Designing an incentive compatible auction that maximizes expected revenue is an intricate task. The single-item case was resolved in a seminal piece of work by Myerson in 1981. Even after 30--40 years of intense research, the problem remains unsolved for settings with two or more items. We overview recent research results that show how tools from deep learning are shaping up to become a powerful tool for the automated design of near-optimal auctions auctions. In this approach, an auction is modeled as a multilayer neural network, with optimal auction design framed as a constrained learning problem that can be addressed with standard machine learning pipelines. Through this approach, it is possible to recover to a high degree of accuracy essentially all known analytically derived solutions for multi-item settings and obtain novel mechanisms for settings in which the optimal mechanism is unknown.

A Comprehensive Survey of Incentive Mechanism for Federated Learning

“…The detailed process of auction can be generalized as bid announcement, bid collection, winner determination, and some other auxiliary procedures such as clearing price, information revealing, etc. In each step, it may contains some specific methods such as first-price payment, second-price payment, etc., which generates a veritable menageries of auctions [38]. Interested readers refer to [37] for deeper technical understanding.…”

AI-driven Prices for Externalities and Sustainability in Production Markets

“…There is some recent work that has adopted a similar agenda, but on markedly different domains and using different approaches. Duetting et al [24] and Shen et al [55] consider the problem of finding optimal prices in revenue-maximizing auctions via deep reinforcement learning, and Cai et al [13] design a policy maker for impression allocation in e-commerce platforms against learning agents. Very recently, Zheng et al [64] consider an abstract socio-economic domain via the lens of deep reinforcement learning for policy making, focusing on taxation as a means of institutional intervention, rather than the adjustment of the prices.…”

“…In terms of the methodology, our work falls broadly into the following three categories: Reward shaping [32,35,51], which refers to adding a term to the extrinsic reward an agent receives from the environment, opponent shaping [37,42,49], which refers to manipulating the opponent (by e.g., sharing rewards, punishments, or adapting your own actions), and automated mechanism design [6,13,24], where an an external agent distributes additional rewards and punishments to promote desirable objectives on a population of artificial learners.…”

Learning Revenue-Maximizing Auctions With Differentiable Matching

“…Another thread of work, sometimes called "differentiable economics", represents auction mechanisms using general parametric function approximators, and attempts to optimize them using gradient descent. [14] introduce several neural network architectures to find revenue-maximizing auctions, including RochetNet, which works for single-agent auctions and is strategyproof by constructions, and RegretNet, which represents auctions as a general neural network and includes a term in the loss function to enforce strategyproofness.…”

“…The theoretical difficulty of devising revenue-maximizing strategyproof auctions has resulted in a number of attempts to approximate them. Recently, [14] presented a method, RegretNet, for learning approximately incentive compatible mechanisms given samples from the bidder valuations. They parameterize the auction as a neural network, and learn to maximize revenue, and maintain strategyproofness, by gradient descent.…”