Logistics ops · Production

DoorDash deploys MLP-gated MoE deep learning model for 20% relative improvement in ETA prediction accuracy

The problem

DoorDash's tree-based ETA models had limited expressiveness — predictions showed less variance than ground truth — and struggled to capture intricate temporal and spatial patterns across a large, varied delivery network as operations scaled.

First attempt

Initial co-training of multitask models caused significant accuracy degradation due to task interference. Enforcing cross-stage consistency via an adjustment to later-stage predictions lowered accuracy. Training Weibull distribution parameters with a log-likelihood loss function produced unreasonable outputs, including negative location parameter values.

Workflow diagram · grounded in source
1
Order triggers ETA computation
trigger
“Time of Arrival = Order Creation Time + Delivery Duration”
2
Embedding feature engineering
ai_action
“we introduced embedding into the ETA prediction model. With embeddings, we can convert sparse variables into dense vector representations”
3
Time series signal collection
ai_action
“we collect time series signals on a minute-level frequency, such as the average order volume per minute over the past 30 minutes”
4
Parallel encoder processing
ai_action
“our MLP-gated MoE model employs three parallel encoders, each serving as an expert in processing different aspects of the input data”
5
MLP decoder combines expert outputs
ai_action
“We bring together the expert opinions from each encoder into a single, rich representation, which is then fed into a multi-layer perceptron to translate the combined insights into an ETA prediction”
6
Multitask sequential training
ai_action
“we've implemented a multitask learning approach to develop our ETA prediction model. This strategy allows us to handle different ETA scenarios together, leading to more consistent and efficient predictions”
7
Probabilistic ETA distribution output
output
“our model doesn't just predict a single ETA value, but rather a distribution of possible arrival times”
Reported outcome

The new MLP-gated MoE architecture delivered a 20% relative improvement in ETA accuracy across large and small orders, long and short distances, and peak and off-peak hours, improving customer trust and operational efficiency.

Reported metrics
ETA prediction accuracy improvement20% relative improvement
Reported stack
MLP-gated MoEDeepNetCrossNettransformer
Source
https://careersatdoordash.com/blog/deep-learning-for-smarter-eta-predictions/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The new MLP-gated MoE architecture delivered a 20% relative improvement in ETA accuracy across large and small orders, long and short distances, and peak and off-peak hours, improving customer trust and operational ef…

What tools did this team use?

MLP-gated MoE, DeepNet, CrossNet, transformer.

What results were reported?

ETA prediction accuracy improvement: 20% relative improvement (source-reported, not independently verified).

What failed first in this deployment?

Initial co-training of multitask models caused significant accuracy degradation due to task interference.

How is this logistics ops AI workflow structured?

Order triggers ETA computation → Embedding feature engineering → Time series signal collection → Parallel encoder processing → MLP decoder combines expert outputs → Multitask sequential training → Probabilistic ETA distribution output.

Related logistics ops cases