Tuning Glovo’s Dispatching Engine at Scale via Optimal Treatment Rules

EasyChair Preprint 13945

Abstract

Glovo operates a three-sided marketplace connecting vendors, customers, and riders to facilitate order delivery. Central to this operation is the dispatching engine, which optimises real-time assignment of orders to riders. The dispatching engine employs a matching cost function to balance customer delivery times and rider efficiency. The matching cost function depends on a number of tunable coefficients that give more or less weight to several important operational metrics. This paper details Glovo's methodology for optimising the coefficients of the matching cost function, focusing on the optimisation pipeline used to determine optimal coefficient configurations and to test them in the real world. We use event-based simulation and multi-objective optimisation to tune these coefficients, ensuring a desirable trade-off between customer experience and operational efficiency.

Additionally, we run switchback experiments and leverage optimal treatment rules to make informed roll-out decisions for new configurations, improving KPIs across diverse markets. Using real data from a Glovo experiment we showcase how our approach based on estimating optimal treatment rules demonstrates significant improvements over naive global roll-out strategies.

Moreover, we report the results of a simulation study comparing the results of estimating the optimal treatment rules using several estimators available in the literature.

Keyphrases: AB Testing, Genetic Algorithms, Optimal Treatment Rules, Optimisation, Switchback Testing, simulation

@booklet{EasyChair:13945,
  author    = {David Masip and Ponç Palau},
  title     = {Tuning Glovo’s Dispatching Engine at Scale via Optimal Treatment Rules},
  howpublished = {EasyChair Preprint 13945},
  year      = {EasyChair, 2024}}