JSM 2021 Online Program

In this talk, we discuss a project from the Graduate Mathematics summer internship program at the National Security Agency (NSA). The students combined methods from reinforcement learning and linear programming to solve a real-world problem. First we introduce the concept of multi-arm bandits (MAB) and explore-exploit algorithms for reinforcement learning. Then we focus on a Baysian MAB strategy called Thompson Sampling. We discuss the implementation of the Thompson Sampling and how to adapt this strategy to handle real-world constraints by inserting a linear programming step between each round of sampling. We give an overview of the advantages of our hybrid approach, such as interpretability for human operators and scalability. Finally we mention how Thompson Sampling can be extended from the Bernoulli payout system to a multinomial.