Conclusions
The goal of this project was to create a neural network visualization
tool that would allow the user to understand the evolutionary and
computational processes involved in training a population of networks
to solve an input-output mapping problem using a genetic algorithm, and
to use this information to more quickly and easily solve problems using
neural networks. We have seen in the course of this report that, in
addition to depicting in graphical form the organization of a single
network and the inheritance relationships among multiple networks in
different generations, the tool developed has provided unexpected insights
into the nature of the search space for a specific problem, and into the
phenomenon of genetic drift. We have found the depiction of
a single network to be useful in the design of networks to solve problems,
the extraction of domain knowledge, and (potentially) optimization of
network topology. Finally, we have seen that this tool can be used to
solve a wide variety of input-output mapping problems, including such
challenging problems as Image Recognition and Breast Cancer
Diagnosis.
As indicated in the section on "Future Work", many opportunities remain
for the enhancement of this tool. Notions of attribute-values, training
set vs. test set, sessions, and many others could all be built into the
program to create a more powerful and useful tool. Alternatively, the
visualization algorithms presented in this report could be used to create
an entirely new tool, possibly for genetic algorithms alone, or perhaps for
neural networks trained in the absence of genetic algorithms. Whatever the
future of this particular tool, I hope it is clear that neural networks are
a powerful tool for solving real-world problems, and that visualization
represents one step toward the achievement of such solutions.
Matt Streeter
March, 2000