Re. Shaffer et Gw. Small, GENETIC ALGORITHMS FOR THE OPTIMIZATION OF PIECEWISE-LINEAR DISCRIMINANTS, Chemometrics and intelligent laboratory systems, 35(1), 1996, pp. 87-104
The application of genetic algorithms (GAs) to the optimization of pie
cewise linear discriminants is described. Piecewise Linear discriminan
t analysis (PLDA) is a supervised pattern recognition technique employ
ed in this work for the automated classification of Fourier transform
infrared (FTIR) remote sensing data. PLDA employs multiple linear disc
riminants to approximate a nonlinear separating surface between data c
ategories defined in a vector space. The key to the successful impleme
ntation of PLDA is the positioning of the individual discriminants tha
t comprise the piecewise linear discriminant. For the remote sensing a
pplication, the discriminant optimization is challenging due to the la
rge number of input variables required and the corresponding tendency
for local optima to occur on the response surface of the optimization.
In this work, three implementations of GAs are configured and evaluat
ed: a binary-coded GA (GAB), a real-coded GA (GAR), and a Simplex-GA h
ybrid (SGA). GA configurations are developed by use of experimental de
sign studies, and piecewise linear discriminants for acetone, methanol
, and sulfur hexafluoride are optimized (trained). The training and pr
ediction classification results indicate that GAs are a viable approac
h for discriminant optimization. On average, the best piecewise linear
discriminant optimized by a GA is observed to classify 11% more analy
te-active patterns correctly in prediction than an unoptimized piecewi
se linear discriminant. Discriminant optimization problems not used in
the experimental design study are employed to test the stability of t
he GA configurations. For these cases, the best piecewise Linear discr
iminant optimized by SGA is shown to classify 19% more analyte-active
patterns correctly in prediction than an unoptimized discriminant. The
se results also demonstrate that the two real number coded GAs (GAR an
d SGA) perform better than the GAB. Real number coded GAs are also obs
erved to execute faster and are simpler to implement.