Time forecasting in logistics software

We started improving from the moment the order was assigned: we calculated the distance that the courier needs to overcome relative to his current position to the restaurant and from the restaurant to the client. To do this, we use an Open Source service that can build routes and calculate time and distance using the transmitted coordinates. Thus, we get two segments:

from the current position of the courier to the restaurant (in logistics it is commonly called the first mile);

and from the restaurant to the client (last mile).

Now we need to calculate the forecast for each participant in the order - the courier, the restaurant, and the customer. This can be done using the formulas:

atVendor (time to arrive at the restaurant) = acceptOrderTime + firstMileTime

pickupTime (time to pick up order from restaurant) = atVendor + cookingTime

atClient (time to be at the client) = pickupTime + lastMileTime

After the calculation, we give forecasts to each participant in the order. Here we can set various coefficients and additional stages of delivery: parking at the restaurant or the client, if the courier is in a car, or the duration of the courier's delay at the client, which depends on the type of payment for the order (cash or card).

How Logistics Software Works

To solve a problem, the algorithm breaks it down into smaller ones, after which it solves them in a certain order or randomly. Kathy Wu, Gilbert W. Winslow Associate Professor of Civil Engineering and Environmental Engineering, and her students augmented this process with a new machine learning algorithm that selects the most useful subproblems instead of solving all of them to improve performance with less computation.

MIT researchers ran a set of subproblems through a neural network they had created and found that this technique accelerated the process of solving subproblems by 1.5 to 2 times. At the same time, it is not known on what principle the neural network selects the best subtasks.

“We don't know why these subtasks are better than others. This will be the direction of our future work. The insights obtained can lead to the development of even better algorithms,” concludes Wu.

Their approach, which they call "delegation learning," can be used in a variety of areas, including route planning for warehouse robots, the researchers say.