How can we transform the everyday technology people use into intelligent,
self-improving systems? For example, how can we perpetually enhance text
messages for managing stress, or personalize explanations in online courses?
Our work explores the use of randomized adaptive experiments that test
alternative actions (e.g. text messages, explanations), aiming to gain
greater statistical confidence about the value of actions, in tandem with
rapidly using this data to give better actions to future users. To help
characterize the problems that arise in statistical analysis of data
collected while trading off exploration and exploitation, we present a
real-world case study of applying the machine learning multi-armed bandit
algorithm TS (Thompson Sampling) to adaptive experiments, providing more
empirical context to issues raised by past work on adaptive clinical trials.
We also conduct field deployments and provide software the community can use
to evaluate statistical tests and algorithms in complex real-world
applications. This helps provide first steps towards integrating two key
approaches: (1) How can we modify statistical tests to better match the
properties of the algorithms that collect data in adaptive experiments? (2)
How can we modify algorithms for adaptive experimentation to be more
"statistically considerate" in being better suited to inference and analysis
of data, while maximizing chances of giving participants useful arms?
Tackling these questions requires multidisciplinary collaboration between
researchers in machine learning, statistics, and the social-behavioral
sciences. This is joint work with Nina Deliu, Sofia Villar, Audrey Durand,
Anna Rafferty and others. The PhD students leading this are Tong Li and Fred
Haochen Song.