I keep losing this article:
http://cacm.acm.org/magazines/2009/11/48443-deep-data-dives-discover-natural-laws/fulltext
Years ago when it was first published, I went through the references and tried to understand how to reproduce the experiment, but got overwhelmed by some of the work.
The idea is that you provide a what I'll call a "vocabulary" - a list of operations - which are then randomly arranged, and that arrangement is scored according to how closely it duplicates a dataset. Call a single operation a gene, and for a large population of random solutions, propagate genes to next generations to converge on higher scoring solutions.
The difficult work is in optimizing convergence. Their work is very impressive, and looks like it's completely free for anyone interested in applying it.
http://www.sciencemag.org/content/suppl/2009/04/02/324.5923.81.DC1/1165893s1.mpg
Are you really sure your architecture is optimized? This is the tool for answering that question.
continued...
No comments:
Post a Comment