problem with these AI approaches is that they're hit and miss somewhat.
I spent approx 3 months working on a tedious dead end (though I think I've seen something I can use later).
I developed and massively multiplayer online game called Stevietopia - a strategy game. The idea was that people would play the game and in the background - as payment - run my client which attaches to my AI distributed processing server...
...and at time I got effectively super PC computing power - I am too poor to be able to afford super PC's but can borrow other peoples
far cheaper
and with this level of processing power, it became clear that the approach I was following was a bit of a dead end - so I'm now going back to 'basics' simpler neural architecture - feedforward nets.
I imagine that racing games that use this sort of cleverness - in general use such networks also.