Colloquium Artificial Intelligence - Dr. D. Bloembergen

Title:
Evolutionary Dynamics of Multi-Agent Learning

Abstract:
Multi-agent problems are typically very complex, and almost naturally require learning and adaptation as part of their solution. Whereas learning in a single agent setting is fairly well understood, the interaction of multiple independent learning agents gives rise to system dynamics that appear non-stationary from the perspective of each individual agent and may exhibit chaotic behavior on the global level. Understanding these dynamics is therefore of utmost importance. In the past decade, evolutionary game theory (EGT) has been used as a framework in which to study these dynamics, based on the seminal insight of Börgers & Sarin in 1997 that the replicator dynamics of EGT match the infinitesimal time limit of simple reinforcement learning algorithms.

In this talk I will provide an overview of these evolutionary dynamics of multi-agent learning, based on my 2015 survey paper on this topic [1]. In addition I will highlight some recent developments in this area which I believe to be relevant to anyone wishing to understand complex multi-agent interactions.

[1] Bloembergen, D., Tuyls, K., Hennes, D., & Kaisers, M. (2015). Evolutionary dynamics of multi-agent learning: A survey. Journal of Artificial Intelligence Research, 53, 659-697.