Ai-Driven Mechanism Design by Weiran Shen & Pingzhong Tang & Song Zuo

Author:Weiran Shen & Pingzhong Tang & Song Zuo [Shen, Weiran & Tang, Pingzhong & Zuo, Song]
Language: eng
Format: epub
Tags: Business & Economics, E-Commerce, General, Computers, Artificial Intelligence, Electronic Commerce, Desktop Applications, Mathematics, Game Theory, Probability & Statistics
ISBN: 9789819792856
Google: Yy_l0AEACAAJ
Publisher: Springer Nature Singapore
Published: 2024-12-24T23:00:00+00:00

Based on the above bidder behavior model, we model the dynamic mechanism design problem as a Markov decision process and use the Monte Carlo tree search algorithm to find a dynamic mechanism that has much better performance than the current one online.

A version of our framework has already been implemented in the online ad auction system in Baidu and has been proven to be able to significantly increase the revenue (see Baidu’s Financial Report of Q1 2018 [8]).

3.2.1.2 Related Works

In the AI community, a recent, interesting line of works aims to tackle the revenue optimization problem from a dynamic learning perspective. Mohri and Medina [43, 45] apply learning algorithms to exploit past auctions and user features. Their algorithms mainly focus on the estimation of the underlying bid distribution and thus rely on the implicit assumption that buyers do not change their behaviors over time. Mohri and Medina [42, 44] aim to optimize revenue with strategic buyers who aim to maximize their cumulative discounted surplus. They give desirable regret bounds to online pricing algorithms. These works assume that there is an underlying value or value distribution for each buyer, and it does not change over time.

Battaglini [11] studies the Markovian consumer model in a long-term contract setting. Their results show that even when the types at different times are highly persistent, the optimal contract is far from a static one. He et al. [25] and Tian et al. [66] also assume that buyers’ behaviors have the Markov property. In particular, Tian et al. [66] focus on buyer behavior predictions and use a truncated Gaussian distribution as the transition probability. Their goal is to find the best static mechanism and restrict their buyer model to be a linear combination of several simple behavior patterns.

Different from these works, our model allows changes in the underlying value distribution and does not rely on simple assumptions of the bidders’ behaviors.

One objective of many existing works in the literature of sponsored search auctions is to improve the revenue of GSP auctions [24, 32, 57, 65]. When designing and analyzing these auctions, most of these works make the standard game-theoretical assumption that advertisers have a single parameter called type that indicates their maximum willingness to pay for a single click-through. When evaluating these auctions, these works also assume that advertisers are rational and will play according to some equilibrium.

While these works shed light on how to design sponsored search auctions in theory, the assumptions they make do not generally hold in the practice of keyword auctions. For example, most advertisers have complex preferences and private information, such as budget constraints [2, 70, 73], multidimensional valuations, and negative externalities [18, 27]. Furthermore, private information such as budget may change dynamically over time, and advertisers may not be able to observe all configuration parameters of the auction.

There are a few exceptions in the literature that take the initiative to design and evaluate sponsored search auctions by getting rid of these assumptions. Ostrovsky and Schwarz [50] conduct large field experiments

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.