Environmental Engineering Reference
In-Depth Information
Neural networks have been used successfully in many domains. However,
one diculty with neural networks is that it is not always easy to understand
what they have learned since the learned information is locked up in internal
neuron states. A related diculty is that there is no way to determine whether
the training set was complete enough to have the network learn the correct
lessons. When faced with a novel situation, the network may choose the wrong
answer. The final di
culty is that controlled adaptation can be good, but too
much learning can cause a previously stable system to fail.
5.4.2 Genetic Algorithms and Programming
Genetic Algorithms are learning techniques based loosely on genetic repro-
duction and survival of the fittest. They use either numeric or symbolic rep-
resentations for the definition of the problem.
Figure
5.10
is a pictorial representation of a genetic algorithm reproduction
cycle. Processing begins by randomly creating the initial pool of genes. Each
gene represents one potential solution and each is evaluated to determine how
well it solves the problem. The best solutions from the current generation are
paired and used to make new genes for the next generation. The bits from
the two parents' genes are mixed to create two new genes. The new genes
enter the gene pool to be evaluated on the next cycle. Since the best genes
are continually being mixed together, the genes in later generations tend to
be better at solving the problem. This cycle continues until a strong solution
is found.
Genetic programs operate in a manner similar to genetic algorithms with
the modification that each gene is actually a small program. Each reproduction
cycle mixes parts of each parent program.
1
0
0
1
1
0
1
0
0
A pair of genes from the
current generation.
0
0
1
0
1
1
0
0
1
Randomly chosen
cut point.
1
0
0 0
1
1
0
0
1
Children genes to add
to the next generation.
0
0
1
1
1
0
1
0
0
Fig. 5.10.
Genetic reproduction cycle
