Artificial Intelligences (A.I.s) are computer programs that are able to adjust their behaviour according to data they see. So A.I.s are able to adjust to the data a human player generates.
There is a number of games which are definitely solved. That means the A.I. plays perfectly:
- Connect Four: A Knowledge-based Approach of Connect-Four. Amsterdam, 1988. Victor Allis.
See also: Solved Game
Computers win always
A second category are games in which A.I.s always win against human players, but they don't have a perfect strategy. Or at least we have not proven that they have a perfect strategy:
- Go on a 5×5 board
- Reversi on a 4×4 board
Update: There are advances on the 19×19 field:
- Nature: Mastering the game of Go with deep neural networks and tree search
- YouTube by nature: The computer that mastered Go
- Google Blog: AlphaGo: using machine learning to master the ancient game of Go
Unspecialized Game A.I.s
The following video is an explanation and demo of software Tom Murphy VII wrote that learns how to play a Nintendo Entertainment System game and then automatically plays it. It's called "learnfun" (for learn function).
You might want to skip to 6:13 for the demo:
Research paper published in SIGBOVIK 2013: "The first level of Super Mario Bros. is easy with lexicographic ordering a and time travel ...after that it gets a little tricky."
There is a follow-up video with Zelda, Punch-Out, Dr. Mario, Contra, Wall Street Kid and Russian Attack:
And a third episode with Super Mario, Gradius, Mega Man 2, Pro Wrestling, Color a Dinosaur, Nintendo Pinball, Cliffhanger, Arkanoid, Double Dare, Ice hockey: