Vol.2, No 3, 2005 pp. 261 - 275
UDC 004.032.26
TOWARD NEURAL NETWORK-BASED
PROFIT OPTIMIZATION
Dragoljub Pokrajac1, Jugoslav Milutinović2,
Zoran Obradović1
1Center for Information Science and
Technology, Temple University
303 Wachman Hall (038-24), 1805 N. Broad St., Philadelphia, PA 19122,
USA
{pokie,zoran}@ist.temple.edu
2The Fox School of Business and Management,
Temple University
1810 N. 13th St., Philadelphia, PA 19122, USA
jmilutin@sbm.temple.edu
A novel technique for profit optimization is proposed. The technique provides
recommendations to management, with an objective of maximizing a profit
function using a neural network-based decision support system. Applicability
of the proposed method is evaluated on simulated precision agriculture
data. The obtained profit increase is compared to the known optimum. Experimental
results suggest that the neural network-based profit optimization techniques
may lead to a significant profit increase; with radial-basis function networks
outperforming multi-layer perceptrons. The quality of provided recommendations
depends on the possibility of learning regression models on training data
from all regions of the attribute space.
Key Words: Neural Networks, Multilayer Perceptron, Radial-Basis
Function Networks, Profit Optimization, Economy, Precision Agriculture
OPTIMIZACIJA PROFITA
ZASNOVANA NA NEURONSKIM MREŽAMA
U ovom radu predlaže se nova tehnika za optimizaciju profita. Tehnika pruža
preporuke menadžmentu, sa ciljom maksimizacije profitne funkcije koristeći
sistem za odlučivanje baziran na neuronskim mrežama. Primenjivost predložene
tehnike je ilustrovana na simuliranim podacima iz domena precizne agrotehnike.
Dobijeni priraštaj profita poredi se sa poznatim optimumom. Eksperimentalni
rezultati sugerišu da primena sistema za optimizaciju profita zasnovanih
na neuronskim mrežama može da dovede do značajnog povećanja profita, pri
čemu radial basis funkcije daju rezultate bolje od višeslojnih perceptrona.
Kvalitet preporuka koji sistem pruža zavisi od mogućnosti učenja regresionog
modela na podacima iz svih oblasti prostora atributa.
