Playing against AI in video games is tricky, because usually AI have been programmed to be efficient, such as taking the best corners and routes in racing games, or being able to micromanage units in RTS games, and basically just being faster and more efficient than human beings.
However could AI be taught to beat other AI? That’s what programmer SethBling has done when he taught a neural network how to play and win at Mario Kart. To teach the neural network how to play Mario Kart, he fed it 15 hours worth of video and also refined some of its behavior, such as taking over the neural network at times to teach it what to do in certain situations.
It seems that his work has paid off as it resulted in the AI being able to win gold by itself in the 50cc Mushroom Cup. Note that the version of Mario Kart he used is the original Mario Kart, where the AI is definitely not as evolved as the more recent versions of the game. However it is still pretty impressive and it’s actually not SethBling’s first time at training AI to play games.
A couple of years ago, the programmer built a neural network called MarI/O where it evolved on its own where it could play Super Mario World by itself.