Dynamic Mechanism Design via AI-Driven Approaches

This chapter explores how the AI-driven mechanism design framework can be applied to optimize dynamic auctions. In real-world applications such as online advertising, the platform sells ad impressions in a repeated fashion, where the buyers’ action space grows exponentially with respect to the number of auction rounds. Consequently, the design space is significantly larger, posing serious challenges in designing desirable auctions. We consider two repeated auction settings in this chapter. In the first setting (Sect. 3.1), we aim to design the so-called cost-per-action auctions, where each advertiser only needs to pay after the user has done specific actions (e.g., purchases). We design a “credit” mechanism, which has both a learning nature and an economic interpretation. Our mechanism is easy to implement and has desirable theoretic guarantees. In the second setting (Sect. 3.2), we aim to tackle the so-called second-order effect, where the advertisers change their strategies in reaction to new mechanisms. We formulate the dynamic mechanism design problem as a Markov decision process and use reinforcement learning techniques to find a solution. This framework has already been adopted by the major Chinese search engine Baidu and was highlighted in Baidu’s Q1 Financial Report of 2018 (Baidu Inc. (2018) first quarter 2018 financial reports).