Published on August 13, 2012 by Dr. Randal S. Olson
4 min READ
If you were to ask a random person what the best example of Artificial Intelligence is out there, what do you think it would be?
Most likely, it would be IBM's Watson.
In a stunning display of knowledge and accuracy, Watson blew away the world Jeopardy champions Ken Jennings and Brad Rutter without blowing a fuse, and ended with Jennings proclaiming, "I for one welcome our new computer overlords."
IBM's Watson represents the current popular approach to AI: that is, spending hundreds of hours hand-coding and fine-tuning a program to perform exceedingly well on a single task. Most people in the field of AI call machines like Watson an expert system because they are designed to be experts at a single task. This approach has been wildly successful lately, producing machines that drive cars and fly UAVs by themselves, beat world chess and Jeopardy champions, and even fool some people into thinking they're human.
However, imagine how hard it would be to hand-code a system that could do everything the human brain is capable of. Do you think that sounds impossible? That's the reason why the field of neuroevolution was born: scientists wanted to harness the creative power of evolution to design the programs that could achieve human-level intelligence.
Neuroevolution, or neuro-evolution, is a form of machine learning that uses evolutionary algorithms to train artificial neural networks. It is useful for applications such as games and robot motor control, where it is easy to measure a network's performance at a task but difficult or impossible to create a syllabus of correct input-output pairs for use with a supervised learning algorithm.
Broadly speaking, the goal of neuroevolution is to evolve an artificial brain with a genetic algorithm to solve a specific task. The artificial brain, oftentimes called the artificial neural network, is designed based off of our understanding of how biological brains work. This video does a great job of explaining artificial neural networks:
As the video mentioned, oftentimes the genetic algorithm starts out with a bunch of random artificial brains. The genetic algorithm then emulates the process of evolution:
This process repeats over and over until the artificial brains master the task. Here's an example of an artificial brain being evolved to walk in a two-legged robot. Notice how the artificial brain does a really bad job of walking at first, but eventually learns walk without falling at all.
Genetic algorithms have been proven to be a creative and powerful designer.
For example, researchers once used a genetic algorithm to design an antenna for one of NASA's satellites. The original antenna took months for engineers to design; cost thousands of dollars per antenna; and didn't even perform as well as NASA had hoped. An entrepreneurial group of researchers at UCSC decided to make an attempt at designing their own version of the antenna with a genetic algorithm, and evolved an antenna that used a single piece of wire that cost next to nothing and performed better than the antenna designed by the engineers.
The same concept applies for evolving artificial brains.
Researchers at UT Austin have evolved artificial brains to control a rocket into space without fins, which is an otherwise extremely difficult problem to engineer. [videos]
Meanwhile, researchers at UCF have evolved artificial brain controllers for two-legged robots that walk and balance all by themselves. [video]
Evolved artificial brains are even being used in video games, such as UT Austin's NERO video game. [video]
There are plenty more examples of "neuroevolution in action" out there; these are just a few choice examples. Neuroevolution has a promising future of designing intelligent algorithms for robot control, vehicle navigation, and many, many, many more applications.
The real advantage of neuroevolution is what it brings to the development of Artificial Intelligence. In the past, computer scientists working on AI would design an algorithm that would exhibit intelligent behavior, then tweak that algorithm's parameters until it exhibited "optimal" intelligent behavior. The AI they designed either worked or it didn't, and oftentimes their results didn't teach us much about how human brains work.
On the other hand, in neuroevolution, scientists can begin to ask questions about the evolution of human-level intelligence:
Indeed, neuroevolution promises to be an insightful field of study, since scientists can not only attempt to create an artificial intelligence, but also hypothesize about how intelligence was created in the first place. (Which is why neuroscientists and biologists are also interested and involved in this field!)