The first set of models are ML models that are trained to estimate how an order will unfold if we offer it to a particular Dasher. Building DeepRed: our dispatch serviceĪt a high level, the dispatch engine is built on two sets of mathematical models. Then we will describe how we leverage our simulation and experimentation platforms to improve our decisions. In the following sections, we will go through our dispatch system’s architecture and how it handles a sample delivery. Both of these methods help us achieve our goals and continue to get 1% better every day. The second stage was to build simulation and experimentation platforms that would allow us to make continual improvements to our dispatch service. First, we built a sophisticated dispatch service that utilizes a number of ML and optimization models to understand the state of the marketplace and make the best possible offers to Dashers to meet the needs of our marketplace. Taking on such a complex problem was a two-stage process. For example, if it's raining and many Dashers use motorbikes we can expect fewer accepted deliveries, which can cause lateness and hurt our ability to complete our goals. We also need to look at conditions like the weather and traffic that can impact delivery times or cause Dashers to refuse orders at higher rates than we would normally expect. These are also situations where it is beneficial to look for batching opportunities where a consumer can get their order faster if a single Dasher is able to pick up multiple orders at the same time. In these undersupply scenarios, we have to make tradeoffs about which orders to pick up now versus later. While we try to make sure there are enough Dashers available to fulfill orders, there may be times when we don’t have enough Dashers to pick up all the orders. The most important one is the supply and demand balance in any given market. There are other marketplace conditions outside of our control that play into our decisions of which Dasher to choose. (This figure appeared in a previous blog post.) Accounting for marketplace conditions Figure 1 The goal of dispatch is to find the best Dasher to pick up each order once it’s ready at the merchant and deliver it to the customer. Another factor is batching, utilizing Dashers as effectively as possible by looking for opportunities where a single Dasher can pick up multiple orders at the same store (or a set of nearby stores). If we dispatch too late, the food will sit too long and could get cold, while the merchants and consumers become upset that the food wasn’t delivered as quickly as possible. If we dispatch a Dasher too early, they will have to wait for the order to be ready. The second factor we look at is ensuring the Dasher will arrive at the right time. We typically want to find a Dasher who is as close as possible to the store to minimize the total travel time. The most important factor is the geographical location of Dashers. To find the best Dasher to deliver an order, we need to consider a number of different factors. We approach each challenge using machine learning and optimization solutions, and use simulation and experimentation methods to build on that performance. Reaching these goals requires overcoming a number of challenges. Deliver orders fast and on time so consumers and merchants are happy with their experience.Propose offers to Dashers as efficiently as possible so they can maximize their earning opportunities.Let’s first define the goals we are trying to achieve when dispatching Dashers. To better understand the dispatch problem we will take a look at the goals DeepRed tries to achieve for each part of our three sided market then we will examine the hurdles we face. In this blog post, we will discuss the details of the dispatch problem, how we used ML and optimization to solve the problem, and how we continuously improve our solution with simulations and experimentation. But how does DeepRed really work and how do we use it to keep the marketplace running smoothly? To power our platform we needed to solve the “dispatch problem”: how to get each order from the store to the customer, via Dashers, as efficiently as possible. DoorDash delivers millions of orders every day with the help of DeepRed, the system at the center of our last-mile logistics platform.
0 Comments
Leave a Reply. |