BCAM Scientific Seminar: How can applied mathematics help optimal decision making? From variational inference to the differential geometry of bandits

Abstract
Optimal decision making is the subject of study in many academic fields, from statistics and computer science to operations research and electrical engineering. Its origins can be traced back to the beginning of the past century, under the mathematical abstraction of multi-armed bandits, with a recent interest sparked by successes in providing super-human performance via reinforcement learning. In this talk, I will discuss how techniques from applied mathematics and physics can help with the understanding of bandits and their applicability: from adopting variational and Monte Carlo inference for arbitrarily complex bandit environments, to using ordinary and stochastic differential equations for studying the dynamics of bandit policies.