Reinforcement Learning for Modeling Platform Economies Under Shock

Abstract

Large-scale platform markets such as Amazon and Uber are playing an increasingly vital and pervasive role in today's economy, especially during periods of economic shock such as during the COVID-19 pandemic. Across many industries, platforms are able to more efficiently facilitate transactions between buyers and sellers, distribute surplus amongst market participants, drive economies of scale, and create value for consumers. However, platforms may also leverage their market power to extract profit at the expense of their users. As such, it is crucial to understand when these scenarios occur by characterizing the drivers of platform economies and systematically studying the complex interactions that take place in these markets.

In this thesis, we use deep Reinforcement Learning (RL) to model platform behavior under various objectives such as revenue, user welfare, and potential regulatory metrics. We then measure the resulting impact on buyers, sellers, and the overall economy. Our contributions stem from the novel idea of using RL to simulate these complex, multi-period platform economies and extend to our extensive empirical simulations and resulting insights. We present a dynamic model in which the platform sets its pricing and matching strategies, while buyers and sellers respond by deciding whether to subscribe and transact on the platform based on the fees and matching quality. We create controlled experiment settings to analyze the effect of various market structures, economic shocks, and platform objectives on the health, diversity, and resilience of the resulting platform economies. We then use these insights to quantify the impact of potential regulatory decisions. As a result, our framework provides a broad foundation for future study and design of platform economies using RL that will hopefully lead to a better understanding of these markets beyond analytical tractability.