I am excited to announce our latest performance improvement: We can now solve a typical last mile delivery problem with 120 stops in 380ms. In this article we show, among other things, how this computing time is composed and which problems take how long.
3 Likes
Amazing work!!
Just out of curiosity, are you planning to use ML in the future to optimize the performance even more?
1 Like
Thank you for your kind words. I’m not quite sure we can use ML to solve these problems even faster. But I can very well imagine using it at some point to learn, with the help of our customers, what good and not so good tours are in order to calibrate our objective functions accordingly. But also here I am not quite sure if ML is the very best or if statistical methods like regression are not sufficient and much better. In any case, I am curious to see what develops here. If you have good literature on this (i.e. route and tour planning specific), please let me know.
I don’t have any good literature unfortunately.
I was in a GIS open source conference, and I think someone from Movit presented their research about using ML for solving routes, I’m not sure I remember correctly, but I think they said they used graphhopper in order to train their model.
In any case, ML is super fast after training as you invest the time in pre-processing, similar to what you do, but differently.