Cuchulainn wrote:outrun wrote:They should have contacted *you*, right?!

Are you dodging a real discussion again?

Anyhoo, if it was my money I would visit Japan to see how they did it. Instead of reinventing the wheel. It would indeed be a good project to manage if we can find a good developer team.

Let me try again. Can you answer this simple question:

*BTW what's the link to ML/AI?? Did I miss something?*The link just seems to add to the impression that this TSP works thanks to ML and AI. At the least it is misleading.Write up or shut up

First, it's not TSP although a set of TSP problems must be solved to test constraint satisfaction of any proposed candidate.

It's a set of TSPs on a set of knapsacks selected to minimize the number of knapsacks. That is they want to minimize the number of buses that serve all the kids going to a particular school subject to both the capacity constraint of the buses and a time-constraint, too. (i.e., eliminate candidate solutions in which a optimized allocation and routing causes a bus to need to pick up the first child at 4:30 AM or risk delivering the students after the start of school).

And even the TSP part probably isn't classic TSP in that: 1) every candidate route is tested for both distance minimization and worst-case time constraint violation; 2) the routing is defined on the road network with stochastic time for traffic and turning direction costs.

The ML part is probably because there's no nice way to mathematically define this nested combinatoric problem (TSP being used to assess optimality and constraint violation on allocated/minimized knapsacks) with messy solution performance estimates (how does one mathematically encode a road graph in which going straight, turning right, or turing left at a node all have different travel costs?)

Instead of MIT, they should have asked UPS which solves this problem every day in setting up the routing of their package delivery drivers. One of UPS's tricks was to bias their algorithms to avoid left turns.