Loading ...
Sorry, an error occurred while loading the content.

Re: [GP] Reply to Koza's recent e-mails concerning the origins of GP

Expand Messages
  • Juergen Schmidhuber
    The following is partly redundant, given my earlier messages. But someone recommended that I respond to Koza lest it appear that I perhaps agreed with him. 1.
    Message 1 of 1 , May 3, 2004
    • 0 Attachment
      The following is partly redundant, given my earlier
      messages. But someone recommended that I respond to
      Koza lest it appear that I perhaps agreed with him.

      1. GP = GA applied to program space.

      2. Program evolution (PE) is older and more general.

      3. Admittedly, if Smith did apply GAs to programs
      in the standard way, as a stand-alone system, not just
      as part of a classifier-like system (where you cannot
      even prove convergence won't get worse over time), then
      Smith's work would qualify as GP and predate Cramer's.
      Koza's reply does not clarify this previously brought-up
      issue, presumably because this would counteract his
      efforts to promote himself as "inventor of GP." Maybe
      someone familiar with Smith's work can do this.

      4. Otherwise, if Cramer used Smith-like crossover
      in a standard GA working on programs, but Smith did
      not, then Cramer was first, except if it's relevant what
      Koza wrote: "Holland discussed using his genetic algorithm
      (with crossover) on assembly code in his 1975 book."
      I did not know this, and do not know to which extent
      Holland discussed GP. Maybe we'd have to consider Holland
      not only as the inventor of GAs but also of the special
      case, GP. (In fact, in my opinion GP does not really
      deserve an extra name as it is just one of many possible
      GA applications, but I did adapt to the fact that the
      term "GP" is now widely used). One obvious question
      would then be: who was the first to implement GP? Does
      Holland and Reitman's 1978 system of if-then-else rules
      qualify as "standard GAs applied to programs"? Does
      Smith's? Or are both systems actually something else,
      say, part of classifier-like systems? Can someone
      clear this up?

      5. Koza mentions that he occasionally cited Cramer
      together with many others and claims: "My practice of
      making bibliographic citations has consistently
      recognized that my work (like everyone else's) builds
      on the good ideas of predecessors".

      This is misleading at best. In places much more visible
      than Koza's rare Cramer citations (e.g. his web sites and
      a recent article in Scientific American) he repeatedly
      claimed he "invented GP" without even mentioning Cramer
      or others, systematically misleading GP novices.

      Koza promoted GP, but no matter how you look at it,
      by no standards he can claim he "invented GP". Clearly
      someone else did.

      Instead of posting long defensive messages, the honest
      thing would be to step forward and say: "True, I did not
      invent GP. But I did a lot to popularize it."

      Juergen



      Meta-GP and Turing-complete GP (1987):
      http://www.idsia.ch/~juergen/diploma.html
      GP with loops etc (1987):
      http://www.idsia.ch/~juergen/geneticprogramming.html
      Overview:
      http://www.idsia.ch/~juergen/gp.html
    Your message has been successfully submitted and would be delivered to recipients shortly.