It has been done, I know, but here is the project I am currently working on (all in Java).
Phase 1: Making a tetris game that io wrappers can easily be made for (done)
Phase 2: Make an IO wrapper so a seperate class can control the game
Phase 3: Make an AI to play the game

A little more detail:
Within the IO wrapper an automatic "reaction time" will be added to make the AIs suffer from human handicaps. The AIs to be added will be an artificial neural network and a genetic system (I havent thought through this one yet). Currently I'm working on the ANN to do this, using the JOONE ANN library. I'm sure I'm leaving out some details that I've thought through but its an overview.

Just curiously, what do you plan on gaining by developing a full fledged ANN to power the game? Wouldn't a finite state system be easier to manage, and play the game much better?

Or are you intentionally trying to simulate the activity of a human? If this were the case would you not add human handicap delays into the NN part of the application? And tech the network when it should and should not delay to produce more precise human-like activity?