Learning connect four. What we see, what the software sees. (via Paris Diderot University)
If you thought designing your own game with relatively no effort was great, then how great would it be having computer learn those games and then beat you at them with relative ease? That’s exactly what computer scientist Lukasz Kaiser, from Paris Diderot University, had in mind with his newly developed software that is able to learn board games just by watching people play. Ok, the software will not actually play your newly designed video-games, but it will learn from you which is still pretty remarkable.
Lukasz’s software is able to learn to play games such as Connect 4, Breakthrough and Pawn Whopping (to name a few) using a decision making engine. Sure there are other software/machine systems that learn to play games, but they usually stick to one game only (think ‘Deep Blue’ chess playing computer) rather than a few varieties. The software (game learning algorithm) was created using OCaml and incorporates multiple logic systems which it uses to learn to play simplistic board-games by processing video of the game being played along with minimal input of the games general rules. The computer Lukasz used to develop the software is no power-house (single-core Intel L9600 processor @ 2.13GHz with 4Gb of ram), but that is the minimum requirement to drive the software to learn a game after analyzing roughly 2 minutes of video and then becoming a formidable component.
As it stands now, the software is only capable of learning simplistic-styled board-games, but Lukasz is planning to upgrade his algorithm to learn much more advanced games such as chess and Stratego in the near future.
Cabe
