Here comes a stick...
I think you need to further define purpose.

There are fundamental limits to what can be achieved in a computer simulation (CS from now on). The input system (mouse, keyboard, etc) limits what is possible to learn as far as actions go, such as shoot, move, communicate - speaking of root learning.

When you simulate a world, you are restricted in what you'll allow to happen in that world. There's a lot of specific programming going into this. That means that whatever solutions people care to come up with, the framework for them has actually really been figured out before. Same thing with threats. The negative side is that the incorporation of new possible actions are not extremely easy to add on, even if developers and users/Soldiers would have direct contact.

Example: A number of insurgents are in a safe house, lacking resources and weaponry. A sweep is long over due and one day a man walks up to the house and says that the Americans know that they have a safe house there. It's a two-story building, so the insurgents sharpen punji sticks, put feces on them (yah, familiar) and fit them under a stair case, which they tamper with to make unstable. Then they leave. As coalition forces sweep the house, they move up the stairs and fall through onto the sticks.

It is such flexibility in situations that is severely limited in CS.

Another problem is scope. Geographically, this is becoming easier as computers become more powerful. On the other hand, an insurgency, based on my little knowledge, seems to depend on social interaction for not only opinion, but also levels of cooperation or resistance, support, etc. Add to this everyone and everything involved in constructing methods for infiltration, evasion, escape - complete networks of people with various levels of resources that need to be able to assess at least local situations and based on that choose to perhaps hide a few days more, or a few hours.
AI is not at a level that can flexibly and realistically handle that today. It does bring us to the next limitation, which is time.

Insurgency is unfolding under a long duration of time, during which insurgents often design their attacks to be suprising. Narrowing of time span restricts this. It also puts a limit on the social processes that are unfolding. Relationships develop bit by bit and if the simulation were to be conducted in a compressed time frame it would be difficult to solve it by having SME:s suggesting reactions to the populace, nevermind how the Soldiers would gauge it.

And with relationships and intelligence comes a feel for interpreting people, and interacting. Emotions, both positive for persuasion and compassion for the father of a child lost, and fear as allied forces move out of an AO and insurgents move in, those feelings don't come along as well when described in ones and zeros.

All the people involved, with AI unable to fill the gap, if you're going to hire them, isn't a role play or FTX better fitted? I think you risk trying to use a tool for which it isn't fit for, where it isn't yet needed because of it's lack of capabilities.

I like what Major Strickland proposed in another thread.

I don't believe this to be God's given truth, but it is a pretty realistic view of what is not possible today. Written for you to tear to shreds, or to see another perspective.

Martin