Problem
Lyft runs a rideshare marketplace where riders, drivers, traffic, events and weather change quickly. Decisions about website copy, messages and marketplace parameters have to follow those changes. Bandits iterate faster and at lower cost than a fixed A/B split, which matters when the best option shifts and there are hundreds of candidate variants.
Example from the article: the rider landing page has several value-proposition texts and main images. The team wants the combination that makes new riders take their first ride, and possibly a different one by time of day or region.
Solution in brief
Lyft uses multi-armed bandit (MAB) algorithms, which shift traffic toward better-performing options while the test is still running. The post covers three uses: UI variants, message and banner selection, and marketplace parameters (incentive budgets, the Wait and Save pickup window, dispatch and pricing parameters). Contextual MAB methods add information such as time of day or local supply and demand.
Questions this case answers
- How to choose among many content variants online while still collecting enough data to learn which one works best
- How to tune a decision parameter in a live two-sided system when conditions shift faster than a fixed test can run
- How to pick the best landing-page text and image combination without running a long fixed A/B split?
- How to choose which message or banner to show a rider or driver when only one slot is available?
Tools and methods
- Multi-armed bandit (MAB) — picks among variants online by balancing exploration and exploitation.
- Thompson sampling with a NIG-Gaussian model — the named algorithm behind the UI optimization tool.
- Contextual and continuous MAB — chooses by context (time, region, demand, driver availability) and handles a continuous choice such as a budget amount.
- Reward function — one number that combines competing goals, such as price and the chance of a future ride request.
Google or a one-time email code. No password.