Tony Xu used to deliver Dominos pizzas from door to door. Today, he is the CEO and co-creator of one of the top AI-driven food delivery apps in the US, DoorDash. Not to mention the poster boy of the American dream.
The platform leverages machine learning to solve the last mile connectivity challenges of local businesses. “DoorDash at the end of the day has to be a phenomenal measurement and data business,”said Tony Xu.
Let’s take a look at the many ways DoorDash uses AI.
DoorDash uses ML algorithms to provide accurate data and mixed-integer programming or reinforcement learning to strengthen their reward maximisation system. “Between the merchants and dashers, it’s the core dispatch problem- how do we match the right set of dashers to the right set of merchants. Between dashers and consumers, it’s about balancing the supply and demand- how do we make sure we have the right number of dashers; how do you make sure we take in the right amount of consumer demand we can take and this really affects the productivity of both the dashers and and the consumers. Between consumers and merchants, it’s the traditional e-commerce application,” said Raghav Ramesh, a machine learning engineer at DoorDash, at an AI conference in 2018.
The platform uses ML to figure out optimal solutions for the Vehicle Routing Problem–to ensure shorter delivery times for consumers, higher pay for dashers and increased income for merchants.
The ML model considers factors such as the food processing time, travel time, delivery location, and the driver’s location to match drivers for delivery. Then, it uses a mixed integration program to determine the cost function. “We make more than ten time point predictions for every delivery and they capture every part of the delivery process,” said Raghav.
Forecasting supply and demand
‘We primarily reformulated the forecasting problem into a regression problem and used gradient boosting through the Microsoft-developed open-source LightGBM framework’, according to DoorDash’s blog. LightGBM allows the platform to train and generate thousands of regional forecasts within a single training run, enabling quick iterations for model development.
Predicting delivery lifecycle
DoorDash real-time, quick-turnaround nature presents extra challenges such as continuous delivery requests, variance in restaurant operations and real-world events such as traffic, bad weather or holidays. The AI creates intelligent logistical systems to navigate the dashers to the right place at the right time. The model considers factors such as the possibility of a worker falling ill and immediately calibrates a restaurant’s offering.
The primary supply and demand measurement considers delivery duration hours, and regional variation is driven by traffic conditions, batching rates, and food preparation times.
DoorDash uses Nvidia graphic processors to speed up the AI training. The engine sorts through the best routes and delivery prices using a complex model. By migrating from single GPUs to Nvidia multiple GPUs, DoorDash has increased its speed by ten times.
The model looks at delivery situation trade-offs to amplify the efficiency of its digital logistics engine. It focuses on market-level metrics to define the state of each market and takes into account metrics like incentivising dashers during high demand periods. For instance, the machine will predict the need for 1000 dasher hours to fulfil the expected demand during a Sunday evening in NY. The machine will further indicate that only 800 dasher hours will be received without incentives, indicating the need for mobilisation for the rest of the 200 hours.
DoorDash recommendation engines have resulted in a 25 percent increase in orders. The algorithm serves up the best selection of restaurants for the consumers depending on their location, quality, past orders, and preferred cuisine. DoorDash also runs targeted promotions for restaurants to create a cyclical system that works in the favour of all three parts of the food delivery equation.