<<

. 87
( 137 .)



>>




K”




st




st
e?




ch


m
fo




lo




“lo
“JF
l?

on




ea nd





f“


f“


f“
ai




of


of




d
em




to


to


to
ph




“m ” a
an
th




l”
t

nt
un




un


un


un




od
ng
y

by




K”
u
b




Co


Co


Co


Co


Co




“fo
“JF
Le
Is


Is




1 0 104 0 1 1 0 1 ... 1 0 ...



Comment Indicator Interaction
Header Variables Variables
Figure 13.5 The comment signature describes the text in the comment.

Team-Fly®
Genetic Algorithms 443


The comment signature is not the genome, but it is related to it. Instead, the
genome is a set of weights corresponding to each variable in the signature
(along with an additional weight called a bias). It is possible to multiply the
weights in the genome times the corresponding fields in the comment signa­
ture to obtain a prediction for the comment being a complaint, as shown in
Figure 13.6. This is the fitness function for a single comment signature. The full
fitness function applies this to all the comment signatures in the training set.
The Genalytics System creates a random population of genomes. These
genomes generally have most of the weights set to low values, and just a few
set to high values. That is, the initial population consists of genomes that are
specialized for the simplest features in the comment signature. Although the
initial population performed very poorly, its use of selection, crossover, and
mutation lead to better and better solutions. After tens of thousands of gener­
ations, the final model was able to classify 85 percent of the records correctly”
enough to speed up the airline™s complaint processing. The chart in Figure 13.7
shows the improvement in the fitness function in succeeding generations.



d”
ge




l”
ea
l”





an
od




ea








K”




“m
st




st
e?




ch


m
fo




lo




“lo
JF
l?

on




nd
f“


f“


f“


f“


<<

. 87
( 137 .)



>>