Dear Candida Ferreira,
>Dear M. Oltean and D. Dumitrescu,
>I transcribe from your paper:
>"The multigenic chromosome was introduced [in GEP] because it can be
>happen that the first symbol in a gene to be a terminal symbol and thus a
>single gene chromosome can not represent a complex expression. As an
>indirect consequence, if the first symbol of a gene is a terminal then
>the rest of the gene is unused."
>To dismiss the multigenic system of gene expression programming this way
is ridiculous.
We don't want to dismiss your multigenic system. We want to show
that this system has some problems. Every method has some advantages and
some disadvantages.
>This is not the reason why multigenic chromosomes were
>introduced in GEP. They were introduced because they are much more
>efficient than unigenic systems. I demonstrate this in my paper
>(http://www.gene-expression-programming.com/gep/webpapers/gep.pdf).
>Besides we know this from nature: multigenic systems are a fundamental
>threshold in evolution as complex organisms require complex genomes. Even
>virus code for more than one gene.
Candida, we are talking here about how to solve problems not about
simulating the nature. You cannot simulate the NATURE on a 900MHz computer
with 128MB RAM.
>So you go on comparing MEP with GEP using a unigenic system. You even
>reproduced three plots from my paper (Figures 2, 6, and 7 in your
>paper). This is copyrighted material and no indication is given about
>permission to reproduce these plots. I used both unigenic and multigenic
>systems in the analysis I made in my paper in order to show the workings
of
>the algorithm and to show the advantages of multigenic systems. For
>instance, if I were to use multigenic systems to analyze the variation of
>success rate with population size, we would get already 100% around
>20-30 individuals. The same is true for the variation of success rate
with
>the number of generations. Again, saturation would have happened around
>30-40 generations. This was the reason why unigenic systems were chosen
>for certain analysis, otherwise we would get no resolution in the
>plots. But this is all clearly stated in my paper. You chose deliberately
to
>ignore this. Any comparisons with GEP should be made using multigenic
systems.
Your multigenic system is hard to use and is not practical. You have to
set manually the number of genes and the linking symbol!
How do you know which are the optimal values? In your paper you manually
set these values only for some problems!
And you also show that if the number of genes is not optimal the
success rate of GEP is very low (Figure 8 in your paper).
How many experiments have you done with BIG data tests? What do you do in
these cases? You set the number of genes to 1 and the linking symbol
to + and run the program, and then set the linking symbol to * and then
run the program, and then set the number of genes to 2 and the linking symbol
to + and then run the program, and then set the linking symbol to * and
run the program and............ ufffffff set the number of genes to 1000
and run the program...........and finally chose the best value?
Multi Expression Programming provides an easy way to maintain multiple
solutions in a single chromosome. You don't have to set manually so many
parameters as in GEP system.
MEP and GEP have the same complexity ONLY IF you know which are the
optimal values for the number of genes and the linking symbol. But you
don't know that so GEP COMPLEXITY IS ONE ORDER OF MAGNITUDE BIGGER THAN
MEP.
And what is all about with this multigenic chromosome?
If you have the expression
E = x^4 + x^3 + x^2 + x
you probably set the number of genes to 4 and the linking symbol to +.
(this is what you have done in Figure 8 in your paper!). In this way you
almost tell to your program what it has to search for! You give partial
information to your program and this is not pure evolution!
And what if the expression does not need operators + or *? You force it to
use these operators or you use a single gene?
Multi Expression Programming is more flexible and do not force expressions
to contain certain symbols. In MEP all needed operators are obtained by
pure evolution and are not provided by the user.
>So, GEP is still the best.
Multi Expression Programming is more flexible than GEP.
MEP requires less human interaction than GEP.
>Candida Ferreira
Best regards,
Mihai Oltean
|