Algorithmic collusion with multi-agent learning
Algorithmic collusion can arise in situations where multiple algorithms that should be competing, learn to work together to the detriment of society. Markets in which pricing algorithms are employed by multiple firms are an example of where this could occur. In order for policymakers to legislate against the use of collusive algorithms, they must know which mechanisms lead to algorithmic collusion. Identifying such mechanisms falls in the broader class of research on the explainability of artificial intelligence. In this talk, I will discuss two different types of collusive algorithms. The first makes use of stochastic gradient ascent methods, while the second employs reinforcement learning. In both cases, it is possible to (partially) identify the mechanisms that lead to collusion.
Science Park 904, room A1.08 and online.