Thursday, February 4, 2010

The Robotic Scientist

Last night, I went to a public lecture entitled The Robotic Scientist: Mining Experimental Data for Scientific Laws, from Cognitive Robots to Computational Biology, by Hod Lipson (Cornell). It was all I had hoped for, and more.

I first heard about this work from a podcast episode. I thought it was cool enough to look up the original paper in the journal Science. Prof. Lipson's lab is using evolution (ie. genetic algorithms) to derive fundamental physical laws from raw data. For example, the figure below shows a double pendulum (a pendulum handing on the end of another pendulum). The graph shows motion-tracking data collected as the double pendulum did its thing. The formula on the right shows an invariant, a fundamental physical quantity that doesn't change, even if the data is doing loopty-loops. In this case, I think the invariant is energy. Invariants are important because they help to point out the general laws that govern how a system works.

In the talk, Lipson told the story of how this project came to be. It all started with using genetic algorithms to evolve simple virtual organisms crawling around in a simulated cyberworld. When they built real versions of these robots, they didn't work if they were to complex. That's because, the more complex the machine, the less realistic their simulations were. So they allowed their system to co-evolve the simulation and the behaviour. That is, they built a 4-legged robot that had 8 actuators and a motion sensor, but no idea what shape it was, or how the moving parts were put together. By performing its own "experiments", the robot soon learned how each of its actuators made it move, and ultimately taught itself to walk (video).

Lipson and his team decided to skip the robot part, and focus these evolutionary computing techniques on the simulation aspect alone. He gave a number of example systems that the program "discovered", like the double pendulum example above. They also solicited data from biology and physics and derived formulas for them. In a number of cases, they showed the formulas to the people who gave them the data, and it turned out to be stuff that was previously invented by famous scientists (for whom the formula was named). It took humans decades to discover what the computer derived in about a day of computing.

Of course, the really interesting question is what to make of the formulas that haven't been named yet? Perhaps their program is showing us something that we've not yet discovered. Hmmm... kinda makes you think.

The talk was part of the outreach program of the Perimeter Institute for Theoretical Physics... you know, your standard small-town theoretical physics think-tank.


No comments:

Post a Comment