Synthetically Controlled Bandits
Vivek Farias3, Ciamac Moallemi2, Tianyi Peng4, Andrew Zheng1
1Operations Research Center, Massachusetts Institute of Technology, United States of America; 2Graduate School of Business, Columbia University, United States of America; 3Sloan School of Management, Massachusetts Institute of Technology, United States of America; 4Department of Aeronautics and Astronautics, Massachusetts Institute of Technology, United States of America
Discussant: Hamsa Bastani (Wharton School, University of Pennsylvania)
We present a dynamic experimental design for settings where the experimental units are coarse (e.g. to mitigate interference). `Region-split' experiments on online platforms are one such setting. Our design, dubbed Synthetically Controlled Thompson Sampling (SCTS), minimizes the cost (i.e. regret) associated with experimentation at no meaningful loss to inferential ability. We provide theoretical guarantees and experiments highlighting the merits of SCTS relative to fixed and switchback designs.