From: Pravin Bhat (pravinb@u.washington.edu)
Date: Wed Dec 01 2004 - 04:31:58 PST
The evolutionary origin of complex features
Richard E. Lenski, Charles Ofria, Robert T. Pennock & Christoph Adam
The paper provides empirical evidence that supports Charles Darwins theory of
evolution by verifying the predictions made by the theory in a simulation
modeled to the approximate the evolutionary process. Special emphasis is given to
verifying the feasibility of complex functions evolving via random mutation and
natural selection.
Strengths:
********************************************************************************
The paper makes a strong case for evolution by replicating its most unlikely
result in a simulation - namely the ability of evolution to evolve really
complex features by starting out with the most basic building blocks. The
EQU function is setup to be extremely fragile in the simulation. It requires
coordination between several simpler functions that could be easily knocked out
by random mutations. However EQU is shown to consistently evolve and survive
in this system using replication, mutation and competition alone.
The work presented in the paper also relevant as a search algorithm. The digital
organisms in the simulation are essentially search engines that are using
massively parallel simulated annealing to find the optimal genome for a given
environment/energy-function. Like simulated annealing the genomes sometimes accept
neutral/deleterious mutations. The probability with which a non-beneficial mutation
is accepted decreases as genomes get closer to the optimal solution since any loss
of major functionality results in a massive loss of energy.
For this search technique to work certain features are required:
- Immense computation power (i.e. the ability to replicate and then proceed in parallel)
- The energy function has to be able to egg the search on towards the optimal solution.
For example, the fact that the authors setup the rewards as an exponential
function of the genome complexity was of immense help to the search.
Flaws:
********************************************************************************
Some of the conclusions drawn in the paper are based on sparse data samples. For
example, the authors argue that the winning genomes that were struck with the
loss of NAND instruction couldnt have emerged as winners without this deleterious
mutation because reversing the mutation nullified their success. This argument
is devoid of any empirical significance as it was based on merely 2 observations.
The experiment is vulnerable to the 'intelligent design' line of argument because
the design of the ancestral genotype is biased towards success. Every ancestor
begins with a NAND instruction which is essential for the system to evolve. Each
ancestor is also given 35 no-op instructions in addition to the 15 instructions
that ensure replication. The 35 no-op instructions essentially increase
survivability during the initial phase by decreasing the probability of initial
mutations knocking out the essential replication instructions. Considering that
neither the no-op instruction nor the genome length was rewarded by the system,
its hard to see how the system could have boot strapped itself in practice without
the help of an intelligent designer or a serendipitous change in the energy function.
Future work:
********************************************************************************
It would be interesting to see a simulation that modeled adversarial evolution
where digital organism derived their energy by hunting other organisms. The
simulation could be used to see the kind of food-chains, attacks and defenses that
would evolve by tweaking the ease with which certain features could co-exist, i.e.
strength and agility would be hard to attain simultaneously. The system would
then either show to build a strong and agile creature by evolving such a creature
or pick the best combination of the feature set that would let a species
survive/dominate in the given environment.
The authors could go a step further to strengthen the case for evolution by
predicting the amount of time and "computational power" that would be required
to evolve the kind of biological complexity we see today assuming that evolution
is just as efficient as their system that evolved EQU. If the predicted time and
computational power matched reality as we know it, the work would provide
stronger evidence for the feasibility of evolution. However this would require
making several simplifying assumptions regarding the amount of "computation power"
encoded by the universe and some rather complex analysis.
This archive was generated by hypermail 2.1.6 : Wed Dec 01 2004 - 04:31:59 PST