EN
Back to the archive

The encyclopedia · Software & IT · Technical decision · 2017–2019

DiDi treated every ride as one move in a long optimization

DiDi frames order dispatching as a Markov decision process, so each match improves response, fulfilment and driver income across the whole day.

DiDi Chuxing

the move

DiDi, China's largest ride-hailing platform, dispatches orders to drivers in real time. Treating each request separately and taking the nearest driver was simple but could not see how a match would affect the next rider or the rest of the day.

The company modeled dispatching as a Markov decision process and used machine learning to optimize assignments across the whole network, so each match is a step toward a better long-run state rather than a locally cheap one.

The winning application improved response rates, fulfilment rates and driver income, and earned the 2019 INFORMS Daniel H. Wagner Prize for implemented operations research.

why it works

  • Independent greedy matches cannot see the next request coming
  • A Markov decision process captures the sequence of trips in a day
  • Machine learning lets the model value a driver's future availability
  • Improving the whole network raises response, fulfilment and income together
the payoffDispatch as a Markov decision processclever

what transfers

Operate the fleet as a linked sequence of decisions, not independent requests; optimizing the long-run state beats greedily minimizing each trip.

what came after

DiDi's dispatching system won the 2019 INFORMS Daniel H. Wagner Prize. The work showed that ride-hailing dispatch could be solved as a large sequential optimization rather than a rule of nearest-first.

references

spotted an error? The archive wants to know.

same kind of clever