Spotlight on Research: Grants Support Machine Learning and Inference Research
Posted: November 19, 2002 at 1:00 am, Last Updated: November 30, -0001 at 12:00 am
By Robin Herron
In the School of Computational Sciences, much of the research is conducted with an eye toward solving real-world problems, and the work being done by Ryszard Michalski and the Machine Learning and Inference Laboratory is no exception. What is more, the National Science Foundation (NSF) and National Security Agency (NSA) are supporting his efforts with three substantial grants.
In the past few years, Michalski, the Planning Research Corporation Professor of Computational Sciences and director of the lab, has focused on the development of two new technologies, non-Darwinian evolutionary computation and inductive databases.
Non-Darwinian evolutionary computation builds on Michalski’s 1997 idea of Learnable Evolution Model (LEM), a new form of evolutionary computation. Unlike traditional evolutionary computation that is based on non-directed evolution, LEM is an evolution directed by a “mind,” or, in this case, artificial intelligence.
In LEM, instead of letting the computer make random mutations and recombinations to create new individuals as in Darwinian evolutionary computation, the system tries to make the “right” ones. Guided by machine learning at each stage, the system creates a theory of “good” direction of evolution, Michalski says. Likening the process to genetic engineering, he gives the example of looking at the genetic makeup of good agricultural crops to see what caused the crops to be good. Then, he says, “The system builds a new population based on the hypothesis as to what makes crops good.”
Working closely with Michalski, Guido Cervone, Michalski’s Ph.D. student, recently developed a new version of a computer program implementing the LEM model. Michalski says, “It worked like magic. It speeded up the evolution hundreds of times more than what was expected. It shortened the evolution so much that reviewers rejected our paper because they couldn’t believe the results.” Eventually, LEM’s results were accepted, and Michalski received a three-year, $345,000 NSF grant to support this research. He also has a patent pending on LEM.
LEM is particularly useful for designing high complexity systems and optimizing functions of a large number of variables. For example, Michalski suggests considering the problem of connecting tubes in a refrigerator so that it has maximum efficiency. “There can be many ways to do it, more combinations than atoms in the universe,” he says. “How do you connect the tubes in the right order? You can do it by intuition, or you can have a computer applying computational intelligence to develop a good design. If you improve the efficiency by even one percent,” Michalski points out, “you have tremendous energy savings worldwide.”
In collaboration with the National Institute of Standards and Technology (NIST), Michalski and colleague Ken Kaufman applied LEM to the problem of optimizing heat exchangers. They used NIST’s simulator to evaluate the capacity of commercial heat exchangers and developed a LEM-based system that proposes new designs by a process of hypothesis generation and instantiation. “It generates descriptions hypothesizing desired properties of a whole class of high-performance designs, and, based on these properties, generates new designs. Eventually, the resulting LEM designs were equal to or better than those designed by humans,” Michalski says.
Michalski adds that LEM provides “many new research opportunities, and they may occupy me for the rest of my life,” but that hasn’t stopped him from pursuing research in other areas.
He’s also working under another three-year NSF grant of nearly $400,000 to do research on “knowledge scouts,” intelligent software agents that scour databases to synthesize knowledge sought by computer users. To explain how they work, Michalski provides the example of planning for a vacation. From information gleaned from previous searches on the Internet, the knowledge scout may know how much you typically spend on a vacation, where and when you like to travel, and what mode of transportation you prefer. The knowledge scout can then offer options that parallel your previous travels.
This research has already found an application to computer user modeling and intrusion detection–topics of importance for computer security and fraud detection. To support this research, NSA provided Michalski and Kaufman with a $300,000 grant.
“One of the biggest problems of today’s world is coping with the flood of information,” Michalski says. “Inductive databases and knowledge scouts aim at helping people cope with the deluge of information to get the right knowledge.”
To promote this area, Michalski chose knowledge mining–the process conducted by knowledge scouts–as the topic for his keynote address at the 2002 International MultiConference in Computer Science, held last June in Las Vegas and attended by more than 1600 people representing 66 countries. Before his conference opening lecture, Michalski was given the 2002 Outstanding Achievement Award by the World Academy of Sciences “in recognition and appreciation of his dedicated and outstanding contribution to the fields of Machine Learning and Artificial Intelligence.”
For more information on this research and the Machine Learning and Inference Laboratory, visit www.mli.gmu.edu.