If artificial intelligence developers like Novamente have their way, virtual worlds like Second Life may soon become training grounds for self-directed artificial intelligence applications which learn by interacting with their environments—and with the virtual people populating them. And maybe users will be able to own their own AI-powered companions.
Although artificial intelligence research dates back decades and famously split into “hard” (rule-based) and “soft” (neural networking) camps, early predictions that intelligent machinery and robots would be improving the standard of living around the world by the end of the 20th century proved to be over-ambitious. Although artificial intelligence applications in well-defined spheres have had some notable successes (such as medical diagnostics and controlling video game opponents), one area where artificial intelligence applications have struggled is interacting with the real world. It turns out that the biological mechanisms and neurological processes by which humans (and animals, and insects, and even plants) perceive and interact with their environment are enormously complex, and enormously difficult for scientists to emulate in non-organic hardware in real time. In other words, it’s hard to make robot or computers see and understand the world the way a human or animal might.
Developers like Novamente aim to advance artificial intelligence applications by taking the real world out of the equation: and virtual worlds like Second Life offer just the means to do that. Instead of having to develop (say) an artificial vision system which continually take snapshots of the real world, attempts to identify objects, people, and even expressions, an artificial intelligence in a virtual world doesn’t have to worry about comprehending its environment: the virtual world is effectively a database it can query and interact with, and the capabilities of the world – gravity, lighting, physical simulation – are already known.
Novamente has developed a Cognition Engine which operates as a “brain” behind an artificial intelligence: it’s a combination of hard-coded behaviors, but is also capable of using reasoning to come up with original ways of meeting its goals. Among Novamente’s first applications for its Cognition Engine in virtual worlds will be virtual pets, which the company expects to debut in 2008 (PDF). Intially, the company is looking at simulations of dogs and monkeys, but hopes to branch out into other animals. Pets are set up with pre-defined goals, like seeking food, getting exercise, seeking “novelty” or new experiences, avoiding danger, receiving rewards and love from their owners and others. Pets will also be able to find their own paths through environments, make variety of inferences, and “learn” using statistical and probabilistic methods. Pets may even exhibit group behaviors, and learn things from each other. The pets would be “embodied” as avatars within the virtual world, complete with their own body language and virtual physiology with which they interact with the environment and other avatars.
Novamente plans to unveil a teachable “virtual companion” in Second Life at the Virtual Worlds Conference in San Jose this October, in conjunction with leading SL developer The Electric Sheep Company. No word if pet licenses will be required.
