I think this would be a really tough problem, harder than for example the well-known travelling salesman problem. A merger of locally optimal solutions for clusters of cities would usually not give an optimal solution for all cities. Each tile with a road on it can be thought as a separate node. Certainly good approximations exist with reasonable run times. Do you have a special interest for the topic?
I also suspect it is a difficult problem! The fact that each road is a node is definitely part of the solution. The hex distances also complicate it, and the fact that the city square itself doesn't really count towards the cost (ie, two adjacent cities have 0 cost, but 1 distance).
My interest is idle speculation, mainly. I learned these topics in university and it is fun to use them when I can. I doubt such an algorithm will be implemented, as the REAL tricky part is planning future city locations. I really doubt the AI would rip up old roads to re-optimize its network for each new city.
Why sure, I have this P=NP solution to the travelling salesman problem on my desk in front of me, would you like me to PM it to you?
Nobody said it had to be polynominal time!

This isn't exactly traveling salesman, though, and it is possible that it doesn't fall into the NP class (as traveling salesman does). In traveling salesman, you can only visit each city once. This is more like a graph connectivity problem, which I don't remember being NP (although it has been a very long time since I learned those concepts, so I might be wrong).