Jul 15 2014
Jeff Clune and his students are not only getting robots to think, but to think creatively.
Jeff Clune, a UW associate professor of computer science, and Jingyu Li, a recent Laramie High School graduate, pose with a copy of the paper they wrote that was published in the Proceedings of the Genetic and Evolutionary Computation Conference. Clune presented the framed copy to her to celebrate her first peer-reviewed publication. (Evolving Artificial Intelligence Lab Photo)
“The long-term goal is to build robots as nimble as natural animals, like hawks and jaguars, and as smart as humans,” says Clune, a University of Wyoming associate professor in the Department of Computer Science. “That is a long way off but, once we can do that, we can send robots, not humans, to put out fires and locate land mines.”
Clune and students in his Evolving Artificial Intelligence Lab wrote a paper, titled “Encouraging Creative Thinking in Robots Improves Their Ability to Solve Challenging Problem,” that was accepted in Proceedings of the Genetic and Evolutionary Computation Conference, a peer-reviewed publication. The paper was published July 12.
Jingyu Li, the paper’s lead writer, was a Laramie High School student when she performed the research. She will attend the Massachusetts Institute of Technology on a full scholarship this fall. Jed Storie, who recently earned his UW undergraduate degree in computer science, was a co-writer.
The paper introduces the “Creative Thinking Approach.” Recent research has shown that robots learn better when they are encouraged to try new things, such as rewarding them for performing actions that they have never done before. The Creative Thinking Approach goes further by encouraging robots to “think differently” about problems, Clune says. Technically, this feat is accomplished by rewarding robots that have ideas they have never had before, meaning different patterns in the neurons in their simulated brains.
Like humans, robots can get stuck in a mental rut while trying to solve problems.
“Our research shows that, when robots are encouraged to think differently about problems, they solve more complex problems faster,” Clune says.
The robots Clune and his team experiment with have virtual brains composed of neurons, just like animal brains. They start an evolutionary process whereby smarter robots get to have more children, leading to “survival of the smartest,” Clune says. “Just as natural evolution produces capable, intelligent animals, their computational version of evolution can evolve smarter robots.”
The innovation in this new research was not just to select for effective, problem-solving behaviors, but also to select robots that thought creatively about the problems they were trying to solve. Specifically, the researchers recorded the firing patterns of all the neurons in the robots’ brains, and then rewarded robots that had novel patterns of neural activity.
Li, Storie and Clune researched the effectiveness of the Creative Thinking Approach by putting the robots through a series of tests, including navigating a maze and collecting a series of balls and putting them away. They tested the technique on increasingly challenging versions of these tasks, and found the harder the problem, the more the Creative Thinking Approach helped. They also showed that their approach outperforms previous techniques that reward behavioral diversity only.
Li says she enjoyed conducting the research and creating increasingly difficult problems for the robots to solve.
“It’s really cool and a lot of fun,” Li says. “I want to do more of this in the future.”
Despite being in high school, Li is a full-fledged member of the Evolving AI lab.
“She is so talented that I treat her the same as all of my Ph.D. students,” Clune says. “She’s hard-working, smart and very creative.”
The study also introduced the concept of a “novelty plateau,” a phenomenon in which approaches that encourage behavioral diversity will have a problem if mutations to the current evolving population do not produce novel behaviors. As in psychology, being stuck on a problem is a result of being stuck in a certain way of thinking. New and creative thinking is needed to solve such problems.
If artificial intelligence were advanced enough today, Clune says human lives could have been spared in a number of recent disasters. He referenced the Fukushima nuclear disaster after an earthquake caused a tsunami to hit the Japanese coast in March 2011; the 19 firefighters who perished in a fire in Yarnell, Ariz., in June 2013; and the ongoing tragedy of Cambodian children losing limbs as a result of detonating land mines.
“Artificial intelligence is at a tipping point,” Clune says. “It’s going from something that was always going to be the next big thing to something that is now the next big thing.”