Computer Teaches Itself to Play Games

Computer games were originally designed to be fun for humans, not for algorithms. Programmer Tom Murphy challenged himself to create a computer algorithm that could learn to play (and beat) Super Mario Bros.

During a human-played session, the algorithm observes the player, mapping events to buttons used. This allows the algorithm to complete levels of the games which were not even ‘learned’ during the input session. This is different from normal artificial intelligence algorithms that are trained to do a specific task.

Murphy's program teaches itself to play rather than it being programmed to take certain actions on specific events. Even though the algorithm was made for Super Mario bros, it works well with any older, side-scrolling game.