On stochastic games, reinforcement learning, and platform competition and collusion

Abstract

In this dissertation we formulate a version of Q-learning with bounded experimentation in a setting of stochastic games with bounded memory and show sufficient conditions under which firms learn that charging supracompetitive prices is optimal in the long run. We also show sufficient conditions for these supracompetitive prices to be supported by three types of different strategies known as naive collusion, grim trigger and increasing strategies. Then, we study what is competition and collusion in a static game model of two-sided markets with an outside option. Comparing collusion to competition, we find that in cases of small cross-side externalities, collusion results in decreased normalized net deterministic utilities, reduced market participation and increased price, on both sides of the market. We quantify the effects of different model parameters in the equilibrium quantities and provide a wide range of economic interpretations. Finally, we examine how AI agents using Q-learningengage in tacit collusion in two-sided markets. We show that collusion by these AI-driven agents is feasible under different choice of the model parameters.