[podcast]https://edwardwillett.com/wp-content/uploads/2009/04/the-artificial-scientist.mp3[/podcast]
As I’ve noted before, the very first science column I wrote, ca. 1991, was entitled, “What is a scientist?”
Last year I re-ran that column with minor editing: the answer to the question hadn’t changed in 17 years.
But it may have changed now.
That’s because researchers at Cornell University have created a computer program that can derive fundamental physical laws from raw observational data.
In other words, they’ve created an artificial scientist.
By observing the behavior of a single pendulum, a double pendulum, and a spring-loaded linear oscillator (things you might use in a high school physics classroom), their software figured out some basic laws of physics, previously discovered by Isaac Newton and successors.
Big difference: it took human scientists centuries. The computer did it in a few hours.
This research arose from previous work by Hod Lipson, a professor of mechanical engineering at Cornell, on a self-repairing robot called Starfish.
Starfish knew how to repair itself because it was able to create a mathematical representation of the ways in which its components worked together over time. Technically, that’s called a “dynamical model,” but you could—and Lipson does—also call it a “self-image.”
With that self-image, it could make predictions about itself, and use those predictions to detect and repair damage.
Lipson and doctoral student Michael Schmidt realized that if the robot could create dynamical models from data about itself, it should also be able to create dynamical models from data about the surrounding world.
The only difference: whereas Starfish created a dynamical model of itself using robot pieces, the new algorithm creates models from mathematical pieces: variables, operators, symbols, functions.
In their experiment, the results of which were reported in the journal Science on April 3, Lipson and Schmidt fed motion-capture data of pendulums and oscillators into the algorithm. The algorithm started with a huge set of mathematical building blocks it could combine in various ways to recreate the patterns it discovered in the data. Through a process called symbolic regression (inspired by biological evolution), it compared the various combinations against each other, searching for the ones that were invariant—that didn’t change from one observation to the next.
Lipson uses the pendulum as an example. “When you look at a pendulum…some things go up, some go down,” he says. “But to recognize when something goes up another specific thing always goes down to keep the total sum constant, this is a key to understanding the observations in a deeper sense—such as recognizing the laws of conservation.”
The computer hung on to the mathematical expressions that remained constant and dumped those that weren’t. That left it with expressions that both matched the data set and could predict future behavior, which were further tested to ensure they were meaningful and not based on coincidental patterns in the data.
Finally, a human examined the results to figure out exactly what the expressions described. The researchers found that the algorithm, given data on position and velocity over time, discovered energy laws; for the pendulum, the law of conservation of momentum; and given acceleration, Newton’s second law of motion.
Using a parallel computer with 32 processors, analyzing simple linear motion took just a few minutes. The much more complex double pendulum required 30 to 40 hours. However, that time could be reduced to seven or eight hours by seeding the problem with terms from equations already derived for the simple pendulum: in other words, by allowing the algorithm to make use of knowledge it had already acquired, just like human scientists do.
Lipson and Schmidt want to continue their research by using their algorithm to examine biological systems, which are notoriously difficult to model. They hope it will be able to find invariant mathematical processes in the enormous sets of data researchers collect about biological systems and thus reveal previously unknown fundamental laws.
They’re quick to point out that their work doesn’t mean computers will make scientists obsolete. “A human still needs to pick the appropriate building-blocks and framework, as well as give words and interpretation to laws found by the computer,” says Schmidt.
To which a science fiction writer such as myself can’t help but reply, “For now.”
