From: Stef Sch... (stef@cs.washington.edu)
Date: Wed Dec 01 2004 - 11:33:25 PST
Paper Title: The evolutionary origin of complex features
Authors: Richard E. Lenski, Charles Ofria, Robert T. Pennock,
Christoph Adaml
This paper examines how complex features can arise in digital organisms
through just random mutation and natural selection, and uses these
simulations to postulate how complex features such as our visual and
auditory systems might occur/evolve in nature.
Strengths:
One of the main contributions of this paper is it's empirical support
of Darwin's theory of evolution. Unlike the real world, in this
digital one they have complete knowledge of all ancestors, and how each
offspring evolves from its ancestors. One of the biggest complains of
evolution is that complex features, such as our eyes, are unlikely to
evolve naturally. For example, a random mutation might cause a simple
light sensing cell to develop, but without any nerves to connect it to
the brain, it would be useless, and therefor no more (and possibly
less, due to it's added complexity) fit to pass that trait on. Their
experiments show that the equals function, which is relatively complex
in their digital system, will consistently evolve using only basic
building blocks.
Furthermore they show that a deleterious mutation at one step may
actually enable a complex operation to evolve in a subsequent
generation, thus at least providing some argument as to how complex
features can evolve in spite (and sometime because of) of harmful
random mutations.
Finally, they provide some evidence that intermediate rewards are
necessary. The equals function builds off of several less complex
functions (such as NAND, XOR, NOT, etc.) In experiments where most/all
of these intermediate functions were rewarded the equals function
consistently evolved and became dominant. However, in experiments
where none of these intermediate functions were rewarded, the equals
function never evolved. This lends some credence to the believe that
complex systems evolve from a sequence of simple increases in fitness.
Namely, it would be quite difficult to evolve an eye from nothing, but
a simple light/dark sensing cell would probably be advantageous. This
gives evidence that small advantages are a requisite for the evolution
of complex features.
Flaws:
They used a strange metric of deleting one instruction at a time to
determine how many instructions were needed for the organism to perform
certain functions. Namely, they didn't analyze the how complex
interactions/redundant may affect the computation of equals (indeed,
one organism apparently needed only 17 instructions to calculate
equals, whereas the best believed one requires 19, so clearly there are
some complex interactions taking place).
Secondly, some of the data was a bit lacking. They only analyzed one
sample as a case study, and in their analysis they didn't look at the
evolutionary steps leading up to the evolution of equals they merely
looked at the previous step. They stated that it evolved from several
different methods, but it would be interesting to see exactly what was
requisite and what was not for equals to evolve.
Finally, they should have analyzed why the equals function didn't arise
in the experiments where it didn't. Given sufficient time, one would
expect this system to evolve the equals function since it is so
beneficial. If the population got stuck in some local optimum, such
that the transition to equals would require too many deleterious steps,
that would be an interesting result and have some potential
ramifications for how we understand the set of species in the world
today. This would be an interesting study, and might shed light onto
why humans only see the visible spectrum, and not infra-red or
ultraviolet.
Relation to Search:
This is clearly a simulated annealing type search problem. The
objective function is the weighted number of functions each organism
calculates, further weighted by the program's complexity. The space is
the space of all programs using these 26 instructions. The operators
are copying with mutations. The start state is the grid initialized
with programs that can only copy themselves, and the goal states is a
program that can achieve the maximum objective function (i.e. one that
can calculate all of the simple functions).
It is slightly different than standard search, in that each of the
programs is competing against the others instead of exploring the space
with others.
Open research questions:
One of the key open questions posed at the end of this is that of
sexual reproduction. It is generally assumed that sexual reproduction
helps, and that it would allow beneficial features to be shared. My
question is, would this benefit the ability to generate equals, or
would it encourage the population to get stuck in some local optimum,
since a single beneficial feature might propagate through the
population, but the changes from it to equals may be too great for such
descendants to survive.
Finally, is this how nature actually works? The authors give some
justification as to why they believe this is at least similar to how
nature works, but this is an open question. Unfortunately, these
questions take an incredibly long time to test in nature, so we'll
probably never know.
This archive was generated by hypermail 2.1.6 : Wed Dec 01 2004 - 11:32:26 PST