In news that might be a little scary for fans of classic film War Games, academics of the Association for Computational Linguistics have taught a computer to learn – by playing games.
In the beginning, the computer had “virtually no prior knowledge” about what it was being asked to do – play a simple game of Civilization. Instead, it had been given a list of actions that it could take – clicking the mouse or moving the cursor. It could access all of the on-screen information, and could tell whether or not it had won.
…but that was about it.
The computer was – understandably – “almost totally random” in its actions, initially. As it played for longer, however, the computer started to recognise different words appearing on the screen following different actions.
The researchers gave the computer a copy of a manual for the game – written in a language it did not understand. As it started to recognise the words appearing on-screen, the computer started to look for instances of those words in the instructions it had been given.
Then, the computer looked at the text to find “associated words”, and piece together actions which those words may correspond to. Some of these actions lead to bad results, so the computer discarded that particular combination.
Others lead to good results, and were given higher priority in the computer’s strategy.
Over time, the computer started showing positive results when it played Civilization. An un-trained machine won 46% of games. Once it had been taught with the manual, the computer won 79%. It’s not yet known whether more time with the manual would lead to an even higher winning ratio.
Eugene Charniak, University Professor of Computer Science at Brown University says:
If you’d asked me beforehand if I thought we could do this yet, I’d have said no.
You are building something where you have very little information about the domain, but you get clues from the domain itself.
These “meaning-inferring algorithms” are now being adapted to work with robotic systems, says [surl=http://web.mit.edu/newsoffice/2011/language-from-games-0712.html]MIT News[/surl].