Multi-objective Mechanism Design via AI-Driven Approaches

This chapter explores how to apply the AI-driven mechanism design framework to balance multiple design objectives. In real-world applications, the platform may care about multiple objectives, such as revenue and welfare. Unfortunately, in general, no mechanism can optimize these objectives simultaneously. In this chapter, we aim to balance these different objectives using AI approaches. Section 4.1 concentrates on the single-parameter setting, where each buyer’s type can be fully characterized by a single parameter. We propose a class of parametric mechanisms. We show that any mechanism in our mechanism class can guarantee an approximation ratio of the optimal one, and the ratio is always better than 2. Section 4.2 aims to design reserve prices for online ad auctions, to trade off between the revenue and the match rate (the overall probability of a sale) or the welfare. We use a machine learning model that takes as input the context information from the auction and gives a reserve price as output. We come up with a loss function that is convex and easy to interpret. Our experiments show that our method Pareto-dominates previous approaches.