DoorDash engineering team describes how they replaced independent supply and assignment systems with a unified joint optimization framework. The new system uses causal ML to estimate the heterogeneous treatment effects of Dasher peak pay incentives and delivery batch rates on delivery duration and cost. Key technical choices include: formulating the problem as a constrained integer program over discretized decision variables, using R-learner with LightGBM as the base causal model, designing low-cost RCT experiments (small-exposure A/B and multi-arm switchback) to gather unbiased training data, and applying a softplus constraint to enforce non-negative HTE predictions. The result is a two-stage proactive/reactive optimization system validated via A/B tests to improve delivery quality while reducing total cost.
Sort: