|
|
|
|
|
|
by debatem1
3402 days ago
|
|
|
It's worth pointing out that Uber had their service working more than a year ago, so while the folks you mention might help take Uber to the N+1th order of optimization they are clearly not necessary to reach the Nth. I suspect this is what the GP is getting at: reaching the levels of optimization on this problem which are required to launch the service is comparatively easy. Going from there to optimality is extremely hard, but may not be necessary from a business or end user point of view. |
|
|
Yes, service can obviously work with subpar algorithms but to really succeed at pricing it as cheaply as possible it requires practically an ability to successfully predict the whole day and then optimizing on the NP-hard problem of that whole day. Maybe sampling a million day variants while routing to make a single decision (which driver should pickup the next request).
Of course brute force greedy algorithms work but they can be, on a hundred vehicle scale, 30% away from the optimum cost.
I've been downvoted to oblivion so I cannot longer keep participating in this discussion (HN will shadowban me).