Uber Nairobi Ambulance Perambulation Challenge
$6,000 USD
Can you use ML to create an optimised ambulance deployment strategy in Nairobi?
1029 data scientists enrolled, 331 on the leaderboard
ConstructionTransportationHealthPredictionStructuredLocation
Kenya
17 September 2020—24 January 2021
130 days
Gradient Descent Solution
published 20 Feb 2021, 14:09

First of all, I'd like to congratulate the winners for their painstaking efforts in making it this far, same goes to every other participant.

From the winning approaches that I've read so far, I noticed that, to my greatest surprise, gradient descent gave the optimal performance. I tried using the different locations as classes in a classification model, and it turned out not to be a reasonable solution. I'd like to know how a gradient descent model could be structured for this particular problem. Were the different location points used as the output in a multi-output regression problem or possible something like it.

I'm interested in seeing how a model could work with this problem and the intuition behind it.

A custom loss function was created by @JohnoWhitaker that for 6 initial locations of ambulances returns a tensor containing the average minimum distance from each accident to the nearest ambulance, then the gradient was computed for that particular loss, and the ambulance locations were updated. Here's the starter notebook by Johno : https://colab.research.google.com/drive/1SKWhBFAEZvhxmBS9ilKQCzcEp923024z?usp=sharing

There wasn't a typical framework used such as classification/regression models.

If it helps, i framed the problem as a binary classification + some post-processing, the output of the classification model was , for each 3H interval and for each road , the probability of an accident, then i center the nearest ambulance around those high-risk roads. It didn't pay off as i didn't have more time to experiment with it, however, after a first trial, i obtained a better score on the private leaderboard than a fixed set of ambulance locations obtained from clustering (kmeans).