Issue: 09/27/06

2006 University Scholar Derong Liu

Can a computer learn from its mistakes?

[print version] [email article]

09/27/06

Paul Francuch

Photo: Roberta Dupuis-Devlin

The University Scholars Program, now in its 22nd year, honors faculty members for superior research and teaching, along with great promise for future achievement. The award provides $10,000 a year for three years.

Can a computer be programmed to learn from its mistakes, just like a person? 

Derong Liu thinks so.

"When I was a child, I imagined machines doing things automatically," said Liu, professor of electrical and computer engineering. 

His curiosity led him to study mechanical engineering as an undergraduate in China.  By the time he entered a master's program at Beijing's Institute of Automation of the Chinese Academy of Sciences, he had switched to a field of electrical engineering called automatic control.

"That required additional study in electronics, circuits and automatic control theory," he recalled.  "I discovered this was something that I'd really like to do as a career."

After finishing his Ph.D. at Notre Dame, Liu worked at the General Motors Tech Center in suburban Detroit and as an assistant professor at the Stevens Institute of Technology in Hoboken, N.J., before joining the UIC faculty in 1999.

As part of his ongoing research, Liu is interested in developing what are called artificial neural networks, which mimic animal nervous systems.

"A neural network is a nonlinear dynamical system.  The human brain is also a nonlinear system,” he explained. 

Liu’s research focuses on adaptive dynamic programming, or approximate dynamic programming. 

“This scheme mimics the human brain's learning capabilities,” he said, adding that the field is in such early stages of study that only a few researchers around the world are working on it.

He and his students are studying new ways to control such devices as computers, car engines and other machines that traditionally work in a straight "a-to-z" manner, considering the multitude of simultaneous factors needed to make an intelligent decision.

"When you learn something, you don't usually use it right away.  You wait until the next time you need it," he said. "The approach we're studying is very similar to this concept. 

“What we eventually can have is a machine which will do something and if it does it good, remembers.  The next time, it will do the same thing.  

“But if it does not do a good job, it also remembers that and next time tries to correct it.  So it basically has a kind of reward or penalty whenever the machine does something."

And how do you reward a computer?

"It's just numbers," Liu said.  

"You tell the computer that it gets a positive number when it does a good job or a negative number when it does a bad job."