AF: Medium: Collaborative Research: Econometric Inference and Algorithmic Learning in Games

点击翻译按钮获取中文摘要

Classical work on economic analysis of the interactions of strategic agents starts with players that have valuations for outcomes, such as items or sets of items they may win in an auction, and analyzes equilibria of the resulting game, where players optimize their strategies to improve their outcomes. To empirically test the prediction of such a theory, one needs to recover valuations of the players. Most econometric methods used to recover valuations rely on the assumption that the game is at a stable equilibrium (known as a Nash equilibrium). It is not surprising that such a framework provides a poor fit to the data in changing or new markets. At the same time, there is a growing theoretical literature in algorithmic game theory that allows one to study games where the game is not at a stable equilibrium. The PIs? program focuses on developing a methodology for inference without relying on the standard notions of the stability of outcomes in dynamically changing environments, such as online auctions. The goal of this project is to develop a theory that allows the researchers to take advantage of new dynamic data sets from electronic markets available on the Internet, and using the findings from the data to further the underlying theory. The results of the project are intended to enable to application and development of Data Science tools for analysis and prediction in non-stable and new market settings. This will affect a broad community of empirical researchers such as market analysts, by allowing them to study economic markets that have previously been considered hard or impossible to analyze.The research program is based on using the theoretical results from algorithmic game theory on game outcomes when players use no-regret learning rules and combine these results with econometric techniques that allow one to estimate the best responses of players from the data using a set of non-parametric estimation techniques. The goal of the program, which PIs initiated in a paper in the ACM Conference on Economics and Computation in 2014, is to combine these approaches to develop a set of analytic tools for empirical analysis of games in non-equilibrium settings. Algorithmic game theory helps one to characterize the properties of outcomes in games (such as approximating factors for revenue and welfare in various cases), where the game is not at a stable equilibrium, assuming players use strategies that guarantee a certain no-regret property in place of the stronger equilibrium best response assumption. The project is aimed at combining the insights from algorithmic game theory with econometric methods to enable the analysis of dynamic markets. The intellectual merit of the project is twofold: (i) providing a methodology for inference in games (i.e., estimation of the payoff functions of players and the distribution of player types) in cases where the players use general classes of learning strategies; (ii) providing tools for the analysis of outcomes in non-equilibrium environments, including the analysis of statistical properties of the outcomes constructed using inferred preferences and types.