Eureqa

» see nutonian.com as it lived, on the wayback machine

~ the goodbye ~

Eureqa did one thing. You gave it a table of measurements and it gave you back a formula that fit them, written out in ordinary algebra instead of buried in the weights of a model. The technique is called symbolic regression, and the way Eureqa did it was to invent random equations, keep whichever ones fit the data best, breed those into the next generation of guesses and repeat until the fit stopped improving. Nothing in that loop knew any physics.

It came out of Cornell's Artificial Intelligence Lab in 2007, written by Michael Schmidt, then a PhD student in computational biology, with his advisor Hod Lipson. In November 2009 they put it up as a free download and it went a long way for a research tool: somewhere above eighty thousand copies, used by people modeling cattle herds and by people modeling the stock market. Lipson's pitch for it was that plenty of fields now had more data than theory, and this was a machine for the gap. In the October 2011 issue of Physical Biology he described a yeast experiment where Eureqa recovered seven equations that were already known, which was the point: you could check it against work that had been done the slow way.

Nutonian, Inc. was built around it to sell the engine to companies rather than laboratories. By the middle of the decade the product had a hosted version and an on-premises version with a trial form, and the front page carried customer logos for BP, Shell, Alcoa and the US Air Force. Whatever else that was, it was a long way from a free download for curious postgraduates.

DataRobot bought Nutonian on 25 May 2017. The press release called Eureqa an A.I.-powered modeling engine and talked about time series analytics; it did not say the word discontinued anywhere, because acquisitions rarely do. The engine went into DataRobot's paid platform and the standalone product quietly stopped being a thing you could obtain. I was a working engineer through all of this and never ran it, which is the ordinary fate of tools that get famous for one demo: the demo travels much further than the download does.

~ the chronicle ~

~ the paper and the pendulum ~

The paper was "Distilling Free-Form Natural Laws from Experimental Data", Schmidt and Lipson, Science volume 324, April 2009, pages 81 to 85. They filmed real oscillators, up to and including a chaotic double pendulum, then fed the motion tracking numbers into the search. What came back, they reported, included Hamiltonians and conservation laws, arrived at with no prior knowledge of kinematics anywhere in the setup. That framing is what made it travel. A machine rediscovering physics from raw measurements is a story anyone can repeat, and for a few years everyone did.

In 2012 Christopher Hillar and Friedrich Sommer published a comment arguing the claim needed qualification, on the grounds that the way the search was structured had Hamilton's equations and Newton's second law already baked into it. The software was still doing something real and useful. It just had more physics in its scaffolding than the headline version admitted.

~ the shell that stayed up ~

nutonian.com kept serving the full Eureqa marketing site for three years after the company behind it was gone, trial links and all, with the logo updated to read "a DataRobot company". The last working capture is 5 August 2020. After that the domain redirected for a while and then broke. Today it is registered on Amazon nameservers with no address published at all. DataRobot's own site mentions neither Nutonian nor Eureqa anywhere.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »