The system, called OWA-RL, does not replace the company’s production assignment optimizer. A store-level policy selects a multiplier that changes the optimizer’s speed-versus-efficiency trade-off, while the optimizer retains feasibility constraints, safeguards, and latency requirements.
The available action space contains five multipliers: 0.8, 0.9, 1.0, 1.1, and 1.2. DoorDash trains a discrete-action Q-network offline on logged marketplace data, joining dispatch runs to later delivery outcomes. Its reward uses customer delivery duration, courier active time, and excess courier active time; rewards are aggregated regionally. The training method includes double Q-learning and a conservative regularizer.
DoorDash reports a global switchback experiment against its production baseline. It says batching increased and courier active and wait time fell, while ASAP delivery duration and 20-minute lateness were statistically unchanged; it identifies 5 p.m. to 9 p.m. as the strongest period. The source does not provide the exact effect sizes in the supplied material, raw experimental data, or an independent evaluation.