The linked paper is worth reading, and an interesting extension to inductive logic programming (ILP), a symbolic-logic-based approach to generalizing from examples.<p>The more general idea of learning from watching someone play isn't <i>that</i> new, but it's usually in more restricted contexts. For example, the system may be preprogrammed with the rules of chess, and then learns how to play chess <i>well</i> from logs of expert play. The work here has some pretty clever representations to allow it to start with a general hypothesis of any board-like game and then narrow down the rules of a particular game by observation.
This is actually very old idea. I remember some game creator program on the Apple II that you could teach games. Probably nothing as sophisticated as chess, but maybe tic-tac-toe and connect four. Unfortunately, this was over twenty years ago, so the name is lost to me at the moment. I'll update if I can find it.
Using chess as an example, and not knowing how the program works, I idly wonder if it would pick up such oddities as: en pessant pawn captures, being unable to castle out of check or to castle across check (e.g. White attempts to castle kingside but Black is attacking f1), the fifty-move rule (I expect it would get threefold repetition), and perhaps even the fact that you can promote pawns to pieces other than queens.