A good number of materials on Benjamin's website.
Renfrew's Xtent model.
The “real” distance between points on a cost surface is affected by the “English Coast length” problem about fractals. Scale can be defined as one man's step, for example. If we study trade tracks, one mule step should be OK.
Also using existing known roads as network edges can work.
We even shoould focus on roads before sites, developing similar predictive models. A path can also be made to go through a steep hillside, if it gets excavated. But normally we are likely to exclude steep zones from our possible road network.