The thing I'm most curious about is how to get this working within more "natural" constraints: i.e., ants in real life can't emit pheromones on their entire path history, only their current location. Is there a way to still successfully run this simulation taking that constraint into consideration?
This is pretty cool, it's one of those "classic" problems in AI/Machine intelligence. I remember working on a slightly less interesting version of this a few (many) years back when I was trying to model a swarm. That was a hardware project, but the basic structure is near identical.
SimAnt (Maxis, 1991) follows this strategy extensively. Ants wander around, dropping different pheromones (food, alarm, ...) and exhibiting surprisingly complex behavior for an RTS of the time.