Can Decentralized Q-learning learn to collude?
Keywords: Decentralized multiagent reinforcement learning, Algorithmic collusion, Best response graphs, Basins of attraction
TL;DR: We quantify the likelihood of learning different equilibria in decentralized multiagent reinforcement learning and use this to study the likelihood of algorithms learning to collude in pricing environments.
Abstract: The possibility of algorithmic collusion between pricing algorithms and the necessary antitrust legislation to regulate against it are hotly debated among academics and policymakers. However, none of the algorithms shown to collude have theoretical convergence guarantees and no theoretical framework exists for characterizing an algorithm's likelihood to collude. In this article, we summarize recent work which provides tools for quantifying the likelihood of collusion for a provably convergent algorithm and applies the results to two simple pricing environments.
Submission Number: 130
Loading