Can Decentralized Q-learning learn to collude?

Date:

During this poster presentation, I cover two recent papers on quantifying the liklihood with which multiagent reinforcement learning algorithms can learn to collude in two simple pricing environments. The first develops the theory and applies it to two-player, two-action games and can be downloaded here. The second applies the theory to a pricing environment with three prices and can be downloaded here.

A summary of both papers is available here.