OK, so this has been burning me up, I love this idea so much... two things to overcome:
1. Source AI Nodes do not have any "desirability" attached to them. They are all equally likely to be picked by the navigation computer. This *could* be easy to implement, but would require low-level hacking of the nav computations. *shudder*
2. Using the heat map would be for the following purposes only: (a) Generating most-likely flow heuristics to salt existing nodes in a map, and (b) allow quick-generation of nodes for a map without them by walking the heat map and auto-laying nodes within an X radius of each step.
Either way, I love it. This is another great way for GE:S to be unique. Our AI is already very unique in that it has a dynamic memory of objects in the world with which the bots can pull from to make their paths. Memory of weapons, ammo, and armor that it comes across to help it better plan its next move given its certain set of conditions.
Another excellent next step would be to combine conditions with node selection. In the heat map generation I can create many different flavors, say: (1) basic movement map, (2) low-health movement map, (3) low-ammo movement map, (4) low-armor, (5) strong weapon... etc etc. Then apply the memory and conditions above with the best-path selection and you get a really smart bot who is reactive to the current state of the world AND himself.