Area awareness?

I think that for AI, area awareness is one of, if not the most important factor. The actual decision maing is important, but accurate decisions cannot be made without accurate information. So, my question is, what is an efficient way to make AI aware of its surroundings? I'm looking for a way to just drop an AI in a level and be able to have it perform with no nodes, precompiled directions, etc. Terrains wouldn't be too much of a problem, but working with complex interior structures like deathmatch levels for games such as UT present a very complex environment.

My first idea, though probably massively inefficient, was a 'true vision' sort of thing. Basically, the screen is divided up into a number of different squares, with smaller squares in the centre and larger towards the edges. Then a series of raycasts determines the average distance of each region (so the middle of the screen, which has more samples of the environment, gives more accurate data). Then some alrogithm (I was thinking a neural net) analyses the relative depth values, factors in the bearing to its location and possibly memory, and decided on a route to take.
This probably would not work...