Bookmark and Share

Alex Bäcker's Wiki / How Biological Evolution Avoids Overfitting
  • If you are citizen of an European Union member nation, you may not use this service unless you are at least 16 years old.

  • Stop wasting time looking for files and revisions. Connect your Gmail, DriveDropbox, and Slack accounts and in less than 2 minutes, Dokkio will automatically organize all your file attachments. Learn more and claim your free account.


How Biological Evolution Avoids Overfitting

Page history last edited by Alex Backer, Ph.D. 11 years, 9 months ago

Most learning systems overfit. Evolution doesn't: it produces genomes that are well adapted to the environment generation after generation, ignoring irrelevant details present only in one generation to provide a lasting fit to enduring features of the environment. How does it do this?


Here, I suggest it achieves this feat through this combination of factors:

1) By constantly testing the fitness of each generation of related but different genotypes, it favors genotype sequences that are robust: fit across a wide variety of continuous variations. This is precisely the opposite of overfitting --a solution that is fit in one situation but fails miserably in others. It is the requirement for a genotype to descend from a continuous string of winners stretching back 3 billion years that makes biological genomes robust.

2) Evolution does not work by developing a single solution to a problem, or a match to an environment. It develops a population of solutions. This allows it to "backtrack" when it hits a dead end by overfitting.

3) Natural selection measures fitness by the number of offspring left behind. Biological organisms are complex enough that in order to leave offspring, an organism needs to perform a number of very different tasks: eating, growing, reproducing. Thus, the fitness measured by evolution is the result of a complex series of features that are hard to achieve by an accidental overfitting.


Alex Backer, Altadena, 10/28/2008. Based on thoughts from 2007 while designing evolutionary algorithms to improve web search.



Comments (0)

You don't have permission to comment on this page.