I've thought about even doing a-- very, granted-- poor man's version of this for the "CPU" opponents in my Atari ST racing game. It's conceptually not very complicated: the "AI" logic performs some action, and if that action contributed to a desired outcome-- getting a frag in Quake III, for instance-- then prioritize that node. If some other action resulted in an undesirable outcome, then demote that node.

On my Atari ST it would have to be extraordinarily simple: for example, what average speed for a car gave the highest probability of getting a high placement in a race? Then it could write that data to disk after each race, and read it back in at the start of all subsequent races.

On a modern computer though, the number of behavior nodes could be quite vast-- and with contemporary processors, the "AI" would "learn" really quickly. The stated results in the article-- that the "AI" logic, reading from its historical records, could beat human players consistently-- doesn't surprise me at all.