The encyclopedia · Software & IT · Technical decision · 2020–2023
Lyft's self-learning matcher kept drivers busier and lifted revenue $30M a year
Lyft switched from matching the closest driver to forecasting each driver's earning value, improving utilization for riders, drivers and the platform.
Lyft
the move
Lyft had always matched the closest available driver to each request. When the pandemic swung demand and supply dramatically, that simple rule stopped working, and managers saw how much system efficiency it was costing.
Instead of a static rule, Lyft built a matching algorithm that forecasts the future value of each free driver — how much that driver will earn over the day — and uses reinforcement learning to assign rides accordingly, improving continuously from live trip data.
The result was genuinely better for everyone: the company generated the equivalent of more than $30 million a year in added revenue, drivers earned more, cancellations fell 3 percent and trips with no available driver fell 13 percent.
why it works
- Nearest-driver matching is greedy and backfires when supply is scarce
- Forecasting driver value captures the long-run effect of a match
- Online reinforcement learning lets the system improve from every trip
- Better utilization raised revenue, driver earnings and ratings together
what transfers
Look-ahead beats greedy: a locally perfect match can starve the next customer, so optimize the value a resource will create rather than the distance it will travel.
what came after
The project made Lyft a 2023 INFORMS Franz Edelman Award finalist. Martin called it the first documented case of a rideshare company using reinforcement learning for matching, and the same self-learning approach now matches most riders.
references
spotted an error? The archive wants to know.