Multi-armed bandits have become a popular alternative to traditional A/B testing for online experimentation at Stitch Fix. We’ve recently decided to extend our experimentation platform to include multi-armed bandits as a first-class feature. This post gives an overview of our experimentation platform architecture, explains some of the theory behind multi-armed bandits, and finally shows how we incorporate them into our platform.
Primer: The Stitch Fix Experimentation Platform
Before getting into the details of multi-armed bandits, you’ll first need to know a little bit about how our experimentation platform works. In our previous post on building a centralized experimentation platform, we explained the #oneway philosophy, and how it makes experimentation both less costly and more impactful. The idea is to have #oneway to run and analyze experiments across the entire business. The same platform is used by front-end engineers, back-end engineers, product managers, and data scientists. And it’s flexible enough to be used for experiments on inventory management and forecasting, warehouse operations, outfit recommendations, marketing, and everything in between. To enable such a wide variety of experiments, we rely on two key concepts: configuration parameters and randomization units.
To continue reading this article, click here.