An interesting post showed up on Slashdot yesterday, about automatic game design.

The basic idea is this: using AI (artificial intelligence) to automatically design games which are fun for real people to play.

The first obvious question is: "what makes a game fun?". Not trivial to answer ... but one theory is that playing games is fun when you're learning (or possibly because you're learning). So a good game is one that has a long (or infinite) learning curve which proceeds at a steady pace.

This definition conveniently answers the second obvious question which is "how can you measure how fun a game is, in the absence of a human?". With learning algorithms which are designed to "train" an AI to play a game, you can measure the speed at which this learning takes place.

The second piece of the puzzle is the variation of the game itself. Given a second set of rules, another AI can vary the conditions of a game which the first AI is learning to play, to try to find the permutation which is the most "fun".

So putting this together, here's what happens:

1. An AI creates endless permutations of a game (by changing speeds, positions of enemies, victory conditions etc.)
2. A second learning AI "plays" each variation
3. The rate at which the AI player improves is measured
4. Games which have learning curves which most closely match "fun-ness" are retained

So what exactly did these crazy AIs come up with?

Here’s a couple of sample games:

Chase the blue:

The whole point of this game is to get to catch the blue thing, twice and fast. It works if you play well and the blue thing doesn’t start way off. The green things move around prettily and teleport when you touch them, but are inconsequential.

Red/green suicide:

This game initially seems impossible. Whatever one does, the score rapidly drops by several points per time step, sometimes reaching levels close to -100. However, after a while, this stops happening, as the red things kill themselves and bring an equal number of green things with them. The player can start incrementing the score, with the goal of bringing it over zero, and usually he is successful in this through just going back and forth on a blue thing or between two blue things if they happen to be placed next to each other.

Race against green:

You have only 28 time steps to reach score 6. Fortunately, the blue things are just lying around, waiting to be picked up and reward you with two points each. Unfortunately, the green things are also out to get (though they are just moving clockwise) the blue things, and when they do, you lose two points! But if you see a green thing in the vicinity, you can easily bump him away to another part of the board without score penalty...

Not likely to set the gaming world on fire, in fact in the author own words:

"They do not represent good game design and are not particularly fun."

But still an interesting beginning, and as the author also says "they are sampled from a limited number of runs of the algorithm described above, and are guaranteed free from human design involvement."

Anyway, hopefully my explanation made some sense, and check out the original article.