From: Vaishnavi Sannidhanam (vaishu@cs.washington.edu)
Date: Wed Dec 01 2004 - 01:19:11 PST
Summary:
--------------
The paper applies the Darwinian evolutionary theory to the realm of digital
organisms. It studies how complex operations evolve from basic operations
and how variations and mutations to the basic operations gradually lead to
the development of complex ones. It explains how replication, variation and
differential fitness results in the development of better digital organisms.
Important Ideas:
----------------------
The idea of using Darwinian evolution for the evolution of digital organisms
is very clever. As the idea is very simple - give high points to beneficial
and complex functions once they have been developed to help sustain them and
to give low points to harmful and useless functions to help weed them out -
is easy to understand and implement.
The paper not only reasons out why the theory of evolution (concept of
rewarding certain functionalities and punishing certain others) in this case
works, but also presents statistical evidence as to how this works, thus
making the study complete.
The paper also chooses asexual reproduction as opposed to sexual
reproduction, which not only reduces the complexity of the algorithm (how to
choose partners to get the best offspring) but also at the same time avoids
situations where selection of wrong partners might cause (like bad
offspring/bad functionalities persisting over generations) .
The paper also shows how sometimes what is termed as a bad mutation (loosing
NAND functionality) at one generation might later on result in something
very good for the next generations to come (EQU).
Flaws:
----------
The paper though briefly mentions the concept of sexual reproduction, it
does not really explain how this would have effected the development of EQU.
It does not do any comparative analysis of asexual vs. sexual reproduction
thus leaving its audience wondering how evolution in each case would have
been different.
Though the paper mentions that evolution of complex functions in most cases
resulted in the loss of a basic functionality for the parent, the paper does
not really explain how the loss of simple functions helped in the evolution
of complex functions. It also leaves us to the question of how the simple
functions should be weighted in order for the complex functions to develop
quickly.
The paper could have also talked about what introduction of high randomness
into the algorithms would do to the evolution of the functions.
Is this problem and solution just another instance of search? What, if
anything, makes this particular search problem different from the usual
search problem?
----------------------------------------------------------------------------
--------------------------------
It can be categorized as a search problem where we start out with minimal
functionality and try to reach our goal state of high functionality.
However, the process of doing the search is a little different as some
mutations might result in regressing back to older search states and taking
a different path to reach the goal state.
This archive was generated by hypermail 2.1.6 : Wed Dec 01 2004 - 01:19:17 PST