The Evolutionary Origin of Complex Features

From: Stephen Friedman (sfriedma@cs.washington.edu)
Date: Tue Nov 30 2004 - 21:06:41 PST

  • Next message: Gaurav Bhaya: "Paper Review"

    The Evolutionary Origin of Complex Features by Richard E. Lenski,
    Charles Ofria, Robert T. Pennock and Christoph Adami

          The authors attempted to show that complex features could be
    developed using the principles of evolution by applying those principles
    to computer programs and evolving a complex logic function, then look at
    the complete traces to see how they evolved, something that isn’t
    possible with current biological evolution due to incomplete
    information. Their first major claim is that complex features generally
    evolve by modifying existing structures and functions. They illustrate
    this showing that their complex function evolved in trials where simpler
    functions are also rewarded, while the complex function did not evolve
    in trials where it was the only rewarded function.
    Another major claim is that are many different solutions to the same
    problem. They base this on the observation that in their different
    trials, many different paths were taken to each of the final organisms
    which demonstrated the best traits. Swapping their assumptions and
    conclusions, at the end of the paper they conclude digital organisms
    allow for studying problems in biology which are can’t be studied in
    organic organisms because of incomplete observability and impractical
    time/space/resource requirements. However, to accept this conclusion,
    one must assume the validity of the hypothesis that the theory they
    demonstrated is indeed true in organic organisms, in order believe that
    showing this organic trend in digital life validates the use of digital
    life.
          One flaw in the paper is that they took a very simplistic approach
    to evaluating the number of “needed” instructions for operations. They
    did this by simply replacing one instruction at a time with nop’s. They
    begin to realize that this is not a good metric, and relax it to a way
    of finding “minimum needed” instructions for a computational trait.
    This is still not correct, for it is easy to imagine a pair of jump
    instructions, one which jumps to the second, and the second jumping back
    to the instruction after the first, which perform no valuable
    computation and could be nop-ed out as a pair, but when nop-ed out
    singly, the second jump appears to be a “needed” instruction. A more
    accurate term for what they found was the susceptibility of EQU to
    deletion via single instruction nop replacements. It is not entirely
    clear to me what valuable information this type of metric provides.
    They use it to claim in an ill-defined qualitative way that EQU is a
    “fragile” function, and that it only sticks around because it is very
    valuable.
          The paper does prompt several questions for further research. The
    first is that put forth in their discussion: what differences occur when
    sexual vs. asexual reproduction is used. Next, their research implied
    there were a large number of potential paths to evolve and organism
    sporting all the computational traits, but is there a real way to get an
    accurate calculation/estimate of the relative number of potential paths
    compared to the total or total reproductively viable? It doesn’t take a
    stretch of the imagination to see that this is just another method of
    search. The space is that of all genome sequences, with the goal of
    finding a viable path through genome sequences that leads to a sequence
    with all useful traits. Thus, knowing abundance/scarcity of valuable
    paths, we could identify problem domains with similar ratios and partial
    solutions where using evolutionary principles would be promising. This
    would also give us a better idea of how well evolution performs at
    finding useful organisms – whether it is a good search heuristic or
    simply taking advantage of dumb luck – and allow it to be compared to
    known search algorithms. Also, they claim that EQU is a more complex
    function, and they assign reward values to functions based on what they
    consider the complexity, but where do those numbers come from. Should
    there be a complexity theory of evolved attributes, and could it be
    applied to organisms to say just how complex an eye is? You never know,
    we may find that a little toe is more complex than an eye.


  • Next message: Gaurav Bhaya: "Paper Review"

    This archive was generated by hypermail 2.1.6 : Tue Nov 30 2004 - 21:06:47 PST