W. B. Langdon and R. Poli
The paper describes 3 popular ways to control bloat. There are
undoubtedly many more, possibly problem specific, ways to control
it. However, in general, bloat will eventually occur with discrete
variable length representations and there is a simple static fitness
function. When trying to match or model (possibly noisy) data, there
seems to be a consensus that bloat should be surpresed, eg to avoid
overfitting. With long term less focused evolution, the situation
seems less clear cut. Perhaps a bloated chromosome can carry
potentially useful information forward to following generations even
if it is not immediately important to the current individual carrying
it? Perhaps a longer representation can accummulate selection neutral
mutations which at some later point will allow it to escape from a
local optima?
You asked some nice questions about natural evolution. While it we
would like to extend the paper to nature there are some stumbling
blocks. Firstly the paper discusses search techniques; Is natural
evoluation a search technique? If so what is it searching for? Less
poetically, its objective function is far from static.
Each DNA base in a cell requires energy (and raw materials) to
construct it. Thus there is some incentive to keep the genetic
material as short as possible (ie we expect some parsimony bias in
nature). [Speculation cells which reproduce rapidly and so double
there genetic material more often, such as bacteria, seem to have much
shorter chromosomes. Similarly virus' have even less. Is this due in
part to the relatively larger cost of their genetic material, ie
increased parsimony bias?] Undoubtedly the situation in nature is far
more complex, but would an analysis based on cost decrease/increase of
each cells genetic material be informative? Perhaps such an analysis
would indicate that most species are close to a neutral point so
either increasing or deacreasing the weight of genetic material in
their cells would reduce the cost overal. If so perhaps no dramatic
change in their genetic material would be anticipated?
Thank you for your questions and I hope the I have gone a little way
to start answering some of them.
Bill
W. B. Langdon, Phone +44 121-414-4791
School of Computer Science, Fax +44 121-414-4281
University of Birmingham,
Birmingham. B15 2TT United Kingdom
http://www.cs.bham.ac.uk/~wbl/