There’s this guy in Switzerland named Julian Togelius. And he set out to do something that I would’ve thought impossible: To teach a computer how to design a fun game.
The idea is this: he created a blank game canvas with a game board and several colored dots, one of which was the player. The other dots had other functions, might or might not move, and behaved differently if they collided with one another (i.e. killing the player, incrementing his score, decrementing his score, etc.) By changing how the colored dots behaved, tons of different ‘games’ could be created. But would they be fun? Togelius had to test each game for "fun-ness" by having a neural-net A.I. "play" the game. If the game could be won immediately by random moves, it was declared too easy. If it couldn’t be won no matter how long the computer plugged away at it, it was too hard. The ‘ideal’ game is one in which the AI couldn’t win at first, but gradually improved as it ‘learned’ to play — Togelius was basing his idea of fun on the notion that learning a new skill is what makes games fun for people.[/quote:d0b85ec95f]