A Grey-Box Approach to Automated Mechanism Design

TL;DR

This paper introduces a grey-box, evolutionary approach to automated mechanism design for double auctions, effectively discovering auction mechanisms that outperform known opponents and possess desirable economic properties.

Contribution

It presents a novel parametrized space for double auctions and an evolutionary search method to optimize auction mechanisms within this space.

Findings

Discovered auction mechanisms that outperform strong opponents in the Market Design Game.

Mechanisms exhibit desirable economic properties when tested in isolation.

The approach effectively reduces the complexity of designing auction mechanisms.

Abstract

Auctions play an important role in electronic commerce, and have been used to solve problems in distributed computing. Automated approaches to designing effective auction mechanisms are helpful in reducing the burden of traditional game theoretic, analytic approaches and in searching through the large space of possible auction mechanisms. This paper presents an approach to automated mechanism design (AMD) in the domain of double auctions. We describe a novel parametrized space of double auctions, and then introduce an evolutionary search method that searches this space of parameters. The approach evaluates auction mechanisms using the framework of the TAC Market Design Game and relates the performance of the markets in that game to their constituent parts using reinforcement learning. Experiments show that the strongest mechanisms we found using this approach not only win the Market Design Game against known, strong opponents, but also exhibit desirable economic properties when they run in isolation.