Man to computer: RTFM

Paul Raven @ 18-07-2011

Via the dashing and debonair Ryan Oakley, researchers at MIT have managed to get a computer to do what most computer users never do, namely Read The Frackin’ Manual. And guess what – the computer’s performance at the task at hand improved hugely! The task in question was… playing Civilisation.

But the task isn’t the point, you see; this is about teaching machines to comprehend input in a linguistic fashion:

The MIT Computer Science and Artificial Intelligence lab has a computer that now plays Civilization all by itself — and it wins nearly 80% of the time. Those are better stats than most of us could brag about, but the real win here is the fact that instruction manuals don’t explain how to win a game, just how to play it.

The results may be game-oriented, but the real purpose for the experiment was to get a computer to do more than process words as data — and to actually process them as language. In this case, the computer read instructions on how to play a rather complex game, then proceeded to not only play that game, but to play it very well.

If you take the same process and replace gaming with something more real-world applicable, like medicine or automotive tech, you could have a computer that’s able to act as more than just a reference tool. A lot more.

If I’m grokking it right, this is the opposite of the approach embodied by IBM’s Watson, which is essentially a search engine on steroids; I’m reminded again of the Chomsky/Norvig debate, and MIT’s approach here looks to be much more in the Chomsky direction. I suspect some sort of synthesis of the two approaches will bring the best results in the long run.

 

Be Sociable, Share!

2 Responses to “Man to computer: RTFM”

  1. Don McArthur says:

    The charming and useful alias ‘fracking’ has been abducted by the fossil fuel industry, hasn’t it? A great loss to online life. In particular, what will irc do?

  2. Rick York says:

    I keep disagreeing (does that make me disagreeable?) with people who say Watson was just a souped up search tool. The signal accomplishment of Watson was its ability to understand natural language.

    One has to remember that the questions asked on Jeopardy are deliberately misleading or obfuscating. By the second day (and that’s important in demonstrating learning), Watson was able to understand them and answer them using his super-duper search engine.