Computers have been beating humans at games for ages, that much is obvious. But a computer scientist from the Universite Paris Diderot in Paris, France has decided to change tack, moving away from complicated data sets, instead creating a vision-based system that can look at a game, learn it, and then play it by learning through "relational structures" instead of long formulae and a database of background knowledge. And it does it very, very well.
The computer is set up to recognise the rows columns, diaggonals and the different pieces in games like Connect 4, Gomoku and Breakthrough, and then programmed with multiple different logic systems, essentially so it can spontaneously generate tactics and formulae on the fly.
"This combination allowed [the system] to generate very short and intuitive formulas in the experiments we performed, and there is strong theoretical evidence that it will generalise to other problems," wrote study author Lukasz Kaiser of Paris Diderot.
"Some of those problems might require hierarchical, structured learning or a form of probabilistic formulas, and in the future we intend to consider such extensions. But already the presented technique signi?cantly improves the state of the art in learning from visual input."
It's an approach that could prove very fruitful for autonomous robots, and at this point it runs in a stock standard single core processor with 4GB of RAM. A supercomputer it ain't.
The system is also able to basically learn enough in two minutes of watching videos of these games being played - whether they're finished or not - to be able to play.
Having said that, we won't be truly satisfied until a computer can do the same with Risk. Or maybe that's just inviting world domination...
Read the entire study report over here.