I'm sure the actual process differs from research group to research group. Here's how one commercially available AI training system claims to work.
http://www.pokersnowie.com/about/technology-training.html
And they go into some of the problems that stem from how they've abstracted the game here: