Genetic Aided Cascade-Correlation

	written by Erik Mayer,  EMAYER@uoft02.utoledo.edu,   Univ. Toledo

     Genetic algorithms are applied to the optimization of the weights in 
the cascade-correlation learning architecture.  In this architecture, the
neural network starts out with a layer of output neurons.  The output
neuron weights are then adjusted to minimize the error in the network.  
A hidden neuron is then added and its weights adjusted so its output 
correlates with the error in the network. It is then connected to the output
neurons and the weights of the output neuron are once again readjusted.  
This process of adding neurons continues until a network with an acceptable 
error is produced.  

     Genetic algorithms are used to find the weights for both the hidden
and output neurons.  We attempt to use the global optimization characteristics
of genetic algorithms to find the global set of weights.  However, while
simple genetic algorithms can find the area of the weight space where there
is a minimum error for the output weights or maximum correlation for the 
hidden layers, they do not converge to the actual minimum or maximum.  We
must find a way to supplement the simple genetic algorithm.

     In our approach, the Genetic aided cascade-correlation, we explore 
the weight space first by using simple genetic algorithms with non-overlapping 
populations and binary encoded weights.  We then use Quickprop to converge to 
the minimum or maximum.  We have applied our algorithm to the two spiral test
with resulting average network sizes of an average of 21.6 hidden nodes.


References

Mayer, E.  "Genetic Aided Cascade-Correlation." COGANN Workshop ICGA-93, 
Champaign-Urbana, Illinois, July, 1993.

Mayer, E.  "Genetic Algorithm Approach to Neural Network Optimization."
Masters Thesis, University of Toledo, Toledo, Ohio, August 1993.

Cios, K.J., Mayer, E., Vary, A., and Kautz, H.  "Neural Networks in 
Analysis of Acousto-ultrasonic Data."  Second International Conference
of Acousto-ultrasonics, Atlanta, Georgia, June 1993.  	

