10 July 2012
THE easiest way to learn a new game is to watch someone else play it – and now computers can do the same. Łukasz Kaiser, who studies logic and games at Paris Diderot University in France, has developed software that learns to play games such as Connect 4 and noughts-and-crosses bywatching videos of humans playing.
As it watches, it uses standard image-processing tools to recognise changes in the separate board squares and pieces of a game, while ignoring extra details like human hands. The videos allow the system to learn the rules by logging what the board looks like when a game has been won, and what count as legal moves.
Having mastered the rules, the software plays the game by examining all possible moves and choosing those it deems most likely to lead to a win. As you would expect, its performance depends on the complexity of the game. Connect 4 has few possibilities, making it very hard to beat the trained computer.
The system is not yet sophisticated enough to understand games where rules for winning are linked to rules for moving, such as in chess where a player loses if their king is unable to make a legal move, but Kaiser is working to tackle that.