IE Decision Systems Engineering Fall ‘20 Seminar Series in Collaboration with LIONS Seminar Series presents “Practical Solutions to Internet Experimentation with High-Dimensional Action Spaces Using Bayesian Optimization,” hosted by Giulia Pedrelli.
Practical Solutions to Internet Experimentation with High-Dimensional Action Spaces Using Bayesian Optimization
Presented by Eytan Bakshy, Facebook
Friday, November 6, 2020
1–2 p.m.
Attend on Zoom
Abstract
Rapid progress in deep reinforcement learning has produced stunning achievements in controlled environments, yet many challenges arise when attempting to apply such methods to real-world problems. Using examples from Facebook, Eytan Bakshy will discuss several problems faced by practitioners who aim to apply RL to their own situations. These include issues with problem specification, safety, off-policy evaluation, deployment, and human factors. Bakshy will present recent work on Bayesian optimization at Facebook which address these concerns, including experimenting in noisy nonstationary environments, multi-objective optimization, combining simulation and real-world experiments, and contextual policy search.
About the speaker
Eytan Bakshy is a principal scientist at Facebook, where he leads the Adaptive Experimentation team. Bakshy’s work focuses on developing robust, general-purpose methods for sequential decision
making under uncertainty, and applying these methods broadly across Facebook and sister companies. His interests include Bayesian optimization, Bayesian machine learning, causal inference and reinforcement learning. Bakshy holds a doctorate in information from the University of Michigan, and a bachelor’s degree in mathematics and computer science from the University of Illinois in Urbana-Champaign.