Governments and regulatory bodies around the world are discussing how to create laws that ensure artificial intelligence (AI) is used fairly and responsibly. One growing concern is that algorithms could learn to work together in ways that harm society, for example, by raising prices. In this project, you will provide a rigorous foundation for these discussions by examining if and how certain types of algorithms can learn such collusive behaviour. You will develop mathematical tools to measure how likely an algorithm is to engage in this kind of coordination by studying how and what it learns and then use these tools to analyze (using theory and simulations) how changes to three key areas can affect an algorithm’s tendency to collude: (1) its learning objective, (2) the representations of its environment, and (3) changes in the algorithmic metagame. By studying these factors, we will improve understanding regarding which design choices lead to harmful coordination and may need to be ...