PS3 Informer

Why Procedural AI is the Next Big Milestone in Gaming

Here's a scene that will be familiar to many current-generation gamers. You are cruising down a city street in your drop top, rap music blaring from the radio. Up ahead, you spot a hot dog stand and pull your car up alongside with a well executed handbrake slide. Your car, like everything else in this world, responds accurately to physics. Jumping out, you demand a succulent and health-restoring hot dog from the vendor. The health benefits of hotdogs aside, so far your mind is blown at how lifelike all of this feels. Then things begin to go south.

Handing your meal over to you, the hot dog seller greets you with his usual cheerfulness: “Here ya go sonny, the finest dog in the city.” Now that you think about it, he always seems to say the exact same thing every time you see him. The vendor then stares at you expectantly as you wolf down your meal. You begin to wish that you could engage this hot dog vendor in some more idle chit chat as you eat, but sadly your communication abilities are limited to bumping into things or shooting them. Hoping to engage the gentleman in conversation the only way you know how, you equip your shotgun and point it at him.

“Arrrrrrrrghhhhhh!” he says, like so many other people you’ve met, and runs away.

How can we create realistic open world games where people that we meet in the street repeat more than the same 3 phrases over and over ad infinitum? Well, there are two main approaches that developers could take. The first approach would involve rounding up hot dog vendors and then recording thousands and thousands of lines of dialog from each, ensuring that the simulated characters you met in the game would always have something unique to say. Of course, this would require the invention of a new kind of storage medium just to contain the millions of lines of spoken dialog that the finished product would contain. This approach is not very practical. Luckily, there is another technology looming on the horizon that could enable game developers to craft truly convincing worlds full of speaking characters: Procedural Artificial Intelligence. This would involve creating complex behavioral algorithms and synthesized speech routines so that characters in the game could literally decide what to say based on how they were feeling that precise moment.

Procedural content generation has already been used effectively by game designers to create incredibly detailed graphical environments using tiny file sizes. Rather than storing all of the necessary textures to render, say, a forest on the game disk, procedural algorithms tell the game what to draw on the fly using numerical instructions. At a basic mathematical level, there isn’t that much difference between a painstakingly hand-drawn texture of tree bark and a short set of numerical instructions that tell the computer how it could draw its own tree bark.

Take for example the first-person shooter game .kkrieger, which was almost entirely created using procedural content generation. Because of its innovative approach, the game weighs in at only 96 kilobytes in size. That’s right, a fully featured first-person shooter with current generation graphical detail can be stored in a file the same size as its own screenshot, pictured below.

This screenshot from .kkrieger contains more bytes of data than the entire game itself.

Procedural artificial intelligence would be based on exactly the same principles used to create textures and graphics, except that algorithms would be used to generate lines of socially-acceptable dialog for game characters. Some designers are already experimenting with procedural approach to artificial intelligence. Right now it is being used to program “flock” routines which allow enemies in popular first person shooters like Killzone 2 to act together as a group convincingly. But much more progress needs to be made before this technology can be used to craft living, breathing NPCs with simulated personalities.

There are two main hurdles that academic researchers need to overcome before game developers can utilize these techniques in their products. The first is that computer scientists need to find ways to more realistically synthesize a human voice. Right now, when you are speaking to a computer-generated voice over the telephone, it is immediately clear that it does not belong to a human being. This effect is similar to the “uncanny valley” feeling we get when looking at a poorly-made 3D model of a person. However, computer scientists are close to getting this technology right, and soon it will be impossible to detect a computer’s robotic speech. In fact, developers will be able to synthesize many diffrerent kinds of voices, male and female, young and old.

However, an even more serious obstacle standing in the way of effective procedural AI is the ability for computers to make meaningful connections between words. Right now, the closest technology we have to this is found in search engines like Google, which create massive databases of words that are semantically related to improve search accuracy. For example, Google knows that “Bridgestone” and “Pirelli” are two words that commonly appear in searches for the word “Tires”. The next step of creating a computer program that can come up with meaningful and grammatically correct sentences is a monumental task for linguists, computer scientists and AI programmers.

The cool thing is that once this technology is perfected, it won’t take long before it starts to appear in our video games. Furthermore, while the thinking involved in crafting procedural code is extremely onerous, the actual amount of physical storage required for something like this is not very weighty. Considering that video games of a decade ago contained only rudimentary AI and physics routines, we wouldn’t be surprised to start seeing games like this in as little as ten years.