Deep reinforcement learning has achieved exciting results, including in video games and robotics. Many domains, including healthcare, education, and customer marketing, involve making decisions that have impact on the real world (a patient's treatment, a student's learning, a company's revenue) and therefore experience (data) is expensive. In such cases it is important to carefully consider what data to collect (balancing exploration and exploitation) and how to maximize the information learned from the data that is collected, in order to make good decisions. I will discuss some of our recent work in tackling these challenges, including more efficient deep reinforcement learning using an approximate Bayesian representation of the uncertainty over the best action, and new approaches to better generalize from existing past data, performing counterfactual reasoning to make higher value decisions in the future.

Bio:

Emma Brunskill is an assistant professor in the computer science department at Stanford University. She was previously an assistant professor in the computer science department at Carnegie Mellon University. She works on how an agent can learn to make good decisions, and is particularly focused on the theoretical and algorithmic challenges that arise in applications where an agent interacts with a person, including education and healthcare. She has received a NSF CAREER award, a Office of Navy Research young investigator award, and a Microsoft Faculty Fellow award. Her lab's work has received multiple best paper nominations and awards, both in artificial intelligence (UAI, RLDM) and in machine learning for education (Educational Data Mining).