Dec 16 2013
In an effort to better understand the evolution of learning, Kai Olav Ellefsen is working to make robots learn more like humans. Specifically, he is studying how more modular brains allow robots to learn new tasks without forgetting what they learned from previous tasks.
The visiting doctoral student -- from the Norwegian University of Science and Technology in Trondheim, Norway -- has spent the fall semester at the University of Wyoming working with Jeff Clune, a UW assistant professor in the Department of Computer Science, and using Mount Moran, UW’s high-performance computing cluster, to create a program to do just that.
“The goal is we want to produce intelligence in robots and machines,” Ellefsen says in the Evolving Artificial Intelligence Laboratory housed in the Engineering Building. “We set up an evolutionary process in the computer that produces a certain amount of intelligence. Machines and robots are far from the level of human learning at this point.”
From the ground up
To accomplish that, Ellefsen had to start from near the bottom of the food chain. For his computer program, he used the example of a mole rat and provided it options for one of its most basic needs -- food. In the program, the mole rat’s survival and ability to reproduce depend on its choices.
“What we wanted to study is how learning happens in the real world and how it evolves, and apply that to the computer,” he says. “By studying it in the computer, it gives us a lot of control over the parameters.”
For example, Ellefsen set up the computer program so that the mole rats had the choice of apples and oranges or poisonous mushrooms during the summer months. The mole rats that ate the fruit -- and avoided the mushrooms -- had a better survival rate.
“If he eats the poisonous mushroom 10 times, he will have a low chance of survival,” he says. “However, if he learns to avoid it after the first time, his chances are much better. This way, the setup of the environment, with poisonous and edible items, encourages evolution to find solutions that learn better.”
To determine whether the rodents could adjust during winter, Ellefsen’s program provided the animals with the choice of nuts or poisonous leaves. Again, the mole rat’s survival rate depended on its food choice. To determine whether the animals could retain previous learned behaviors, Ellefsen had the mole rats cycle through another summer foraging season. He emphasized that food associations learned during winter and summer are completely different.
“Evolution can form a solution that is robust, so that when you learn something one summer, you will know it next summer, too,” Ellefsen says. “It’s survival of the fittest.”
Modular thinking
The evolutionary process started with 400 mole rats and ran 20,000 generations of the animals through the simulated environment. Evolution was optimizing the behavior of the mole rats by tuning an artificial neural network, a computational model of the human brain.
Ellefsen had to use a standard industry model that functions more like a human brain -- and work to encourage the model, through evolution, to use fewer connections as a means to avoid interference in learning.
Much like the human brain is separated into modules that allow a person to perform multiple different functions, Ellefsen hypothesized that those artificial neural networks that focus on short and few connections will function better than those that do not. To investigate this hypothesis, he compared neural networks that are encouraged to have few connections (during evolution), and those that are not encouraged to any specific form of connectivity.
“Connections require energy and resources. To help evolution find good solutions to this problem, we encourage fewer connections,” he says. “Fewer (brain) connections help us avoid interference in the learning of different skills and associations.”
He sees endless opportunities to improve robot functions, and mentions the self-driving car as a type of robot that is closer to reality and can perform a human function. Robots also could be used for rescue missions or to assess natural disasters before humans enter the scene.
Clune agrees
“Instead of asking humans to risk their lives to fight fires, we should be sending robots. We can’t do that currently because robots do not learn tasks very well,” Clune says. “But the research Kai and I are doing will help robots better be able to adapt to new situations and solve new problems. Eventually, they can save lives -- though, of course, we have a long way to go.”
Ellefsen says he had a desire to locate an “interesting lab in the U.S.” to pursue research in artificial intelligence, specifically evolutionary competition. Keith Downing, Ellefsen’s faculty adviser in Norway, knows Clune and told Ellefsen about Clune’s research on artificial intelligence and robotics.
“Keith told me that Kai is a smart, hard-working scientist with a strong background in our field. He also told me that he (Ellefsen) is a smart, funny person,” Clune says. “That made it a no-brainer to invite him to become a visiting Ph.D. student in the Evolving Artificial Intelligence Lab.”
It’s turned out to be a good match. The project combined Ellefsen’s interest in evolutionary learning with Clune’s modularity work.
“In just four short months, Kai has gone from an idea to a set of ground-breaking results that we are currently writing up for publication,” Clune says. “I’m extremely excited about the important results he has worked hard to uncover.”
“He’s (Clune) given me a lot of great ideas,” Ellefsen says. “It’s been a lot of fun to work with him and in his lab.”
Faster processing
And Ellefsen has appreciated being able to use Mount Moran, UW’s high-performance computing cluster nicknamed after a mountain peak in western Wyoming’s Tetons.
“We set up a connection and sent the program to Moran,” he says. “We set it to run for thousands of generations. It takes six hours to a few days.”
Mount Moran enables atmospheric and earth sciences faculty -- who will be able to use the NCAR-Wyoming Supercomputing Center (NWSC) -- to learn what to expect with their software. The cluster provides the opportunity for that group of faculty members to work out issues caused by scaling up parallel algorithms from tens or hundreds of processors to thousands of processors, before moving up to tens of thousands of processors on the NWSC.
The cluster also provides a research resource for any UW research faculty members -- such as bioinformaticists, social scientists, pure mathematicians and theoretical physicists -- who have complex problems or whose research doesn’t fall within the scope of the NWSC.
In addition, UW students are welcome to use the high-performance computing center for their work, which is often in concert with UW faculty members.
Ellefsen and Clune are writing a research paper on their findings, which will serve as part of Ellefsen’s doctoral dissertation.
“By trying to evolve learning ability in a computer, we can see robots in the future that can go into the real world and solve problems; demonstrate behavior that current robots don’t have; be more able to modify behavior; and handle new challenges as they go along,” Ellefsen says. “The world is complex and changing, and we need robots to change with it.”