Paper available: "Training Neural Networks by means of Genetic Algorithms Working on Very Long Chromosomes" by Peter Gravild Korning

[Allan Ove Kjeldberg]

The paper korning.nnga.ps.Z is now available for copying from the 
Neuroprose repository:

"Training Neural Networks by means of Genetic Algorithms Working on very
Long Chromosomes"

Peter Gravild Korning
University of Aarhus
Denmark

The paper addresses the problem of training neural nets by use of a genetic
algorithm, where the GA constitutes a general technique, an alternative to
e.g. back-propagation.

Attempts to do this have failed in the past. This is primarily due to the
fact that the mean square error function known from back-propagation has
been used as fitness function for the genetic algorithm. I have invented a
new fitness function which is simple but very powerfull. Unlike the mean
square error function, it takes into account the holistic nature of the
GA's search. And the results are very promising.


All critique and all comments/suggestions are very wellcome.  (please use
the mail address aragorn at daimi.aau.dk or korning.cbs.dtu.dk)


	Peter GRavild korning




ABSTRACT:
In the neural network/genetic algorithm community, rather limited success
in the training of neural networks by genetic algorithms has been reported.
In a paper by Whitley et al. (1991), he claims that, due to "the multiple
representations problem", genetic algorithms will not effectively be able
to train multilayer perceptrons, whose chromosomal representation of its
weights exceeds 300 bits. In the following paper, by use of a "real-life"
problem, known to be non-trivial, and by comparison with "classic" neural
net training methods, I will try to show, that the modest success of
applying genetic algorithms to the training of perceptrons, is caused not
so much by "the multiple representations problem" as by the fact that
problem-specific knowledge available is often ignored, thus making the
problem unnecessarily tough for the genetic algorithm to solve. Special
success is obtained by the use of a new fitness function, which takes into
account the fact that the search performed by a genetic algorithm is
holistic, and not local as is usually the case when perceptrons are trained
by traditional methods.