At Facebook, decision making under uncertainty is central to optimization and product development. On the Adaptive Experimentation team, the challenges we seek to address are often characterized by the following features:
1) The decision-making / experimentation process is sequential, iterative and goal-driven;
2) Objective functions are expensive-to-evaluate;
3) Parameter spaces are often continuous and complex.
4) Exploration must occur in a “safe” way as to enable live experimentation.
One of Facebook's tools for dealing with these problems is Bayesian optimization, for which we've released a research framework, BoTorch, and an experimental management system, Ax. Our work, however, also uses a broader variety of techniques, drawing from causal inference, bandits and reinforcement learning.
The goal for this workshop is to think about how to integrate the insights around these fields, and to think about how they can be used productively in industry. We are hosting this workshop to bring together distinguished academics and industry researchers to discuss the present and future for a research agenda around better decision-making through adaptive experimentation.