GDC 11: Experimental AI Demos Guide You Through the Uncanny Valley
For all their graphical prowess, games still have a long way to go in the realm of AI. Thankfully, the people working on solutions to this problem are legion. Tuesday’s experimental AI panel showcases the efforts of some of these AI programming pioneers.
First on the docket were the attempts of the Digipen Institute of Technology to create a believable Interactive Human. The “Uncanny Valley” is the designer’s perpetual enemy in this task, and the Digipen folks first thought was to avoid this problem by embracing it — trying to create, say, an Interactive creepy baby doll.
What they ended up with was far less terrifying. A young girl, blond and fresh-faced, appeared onscreen, and the designers explained the process by which they made her more believable and interactive. First up were passive facial movements like breathing, blinking, and subtle, involuntary movements of the head. Following these additions, they layered on emotional cues, turning the girl’s blank stare into a look of worry. Finally, they added story and context to the emotion, and the girl explained in a halting voice that she was looking for “her mommy.” When all these elements combined, it was clear that Digipen had made huge strides towards realism, though they ended with a flourish, animating their adorable little girl shouting Schwarzenegger’s unforgettable “get to the choppa!” dialogue from Predator.
The next two projects on display were both born out of UC Santa Cruz’s computer science PhD program, and both tried to tackle the challenge of iterative, AI-powered design, albeit on a fairly simple basis (Hellgate: London, anyone?). The first program created endless platformer levels, divided by discrete vertical intervals. A designer could then make some intervals permanent, and generate new iterations of the level around them, or instead choose to subdivide the intervals to increase the complexity of their creation.
Another program, “Variations Forever,” was by far the most interesting. Though the particulars are difficult to describe, it was effectively an engine for generating classic game mash-ups. Furnished with a number of simple variables and rule-sets from classic games (Pac-Man, Asteroids, and many more), the program was able to generate an infinite number of simple 2D games. They weren’t always playable, or winnable, but seeing an AI create new games on the fly was a powerful thing.
The panel wrapped up with “Occupancy-Regulated Extension,” a complicated name for a program that replicated the design processes of many familiar games, most notably Diablo. The question it sought to answer was this: “how do we get the good elements of human-powered design into procedurally-generated content?” The answer, it turns out, is to create a set of granular, quality-assured, human-designed chunks, then allow the AI to recombine them at random. Then you kill Mephisto, start up a new game, and hope that you get better drops this time around.