Like a locksmith with a master key, Carlos Guestrin has created a computer algorithm that can do everything from figuring out the best way to detect water contamination to revealing which political blogs do the best job of staying on top of the news.
The Carnegie Mellon University computer science professor’s work landed him on this year’s “Brilliant 10” list created by Popular Science magazine, showcasing some of the nation’s top young researchers.
The key to Mr. Guestrin’s algorithm, which is a recipe that a computer uses to solve complex problems, is that it can assemble the maximum amount of information with the least effort.
Mr. Guestrin, 33, began work on the project four years ago when he was doing research for Intel in Berkeley, Calif. Scientists there wanted to measure the climate variations under the soaring canopies of the redwood forests by installing a network of wireless sensors.
“I said to them, ‘How will you decide where you’re putting the sensors in the forest?’ And they said, ‘Oh, whatever looks good, we’ll use our intuition.’ ”
But Mr. Guestrin thought he could come up with a more logical approach.
“I thought maybe I’d do a project where I’d work with them a month on this, and four years later, we’re into something much bigger and more elaborate and interesting than I thought I could possibly get into.”
In a simplistic sense, Mr. Guestrin’s algorithm figures out the best place to put the first sensor in a complex network, then the best place for the second-best sensor, and so on down the line until there is not much incremental value in adding another sensor.
“If I’ve put out four sensors,” he said, “putting out a fifth one might help me a lot, but if I’ve put out 100 sensors, the 101st will help me less, and if I put out 1,000, the 1,001st will help even less.”
Knowing the minimum number of sensors needed to get optimal information can be especially important when the measurement devices are expensive.
That was a key issue when the federal Environmental Protection Agency wanted to know the best way to detect contamination in an urban water system. In the model his team was working with, Mr. Guestrin said, there were 12,000 possible locations to put sensors, but even the cheapest sensor cost $14,000.
Using his algorithm, he was able to show that officials could get a good answer on water contamination by using just 19 strategically placed sensors.
Mr. Guestrin’s algorithm is part of a broader effort under way in computer science known as machine learning, which he defined as “computer programs that can improve their performance based on experience.” Machine learning programs, in other words, can discover new information on their own.
The scientist grew up in Rio de Janeiro and got his bachelor’s degree in mechanical engineering from the University of Sao Paulo. He attended graduate school at Stanford University, worked for a year at Intel after getting his doctorate in computer science, and then joined the Carnegie Mellon faculty in 2004.
He is now working on long-range projects to develop algorithms that would allow people to make better decisions using information from the Web and to allow computers to do a better job of divvying up tasks to solve large-scale problems.