Open AccessNeural Network Assisted Experimental Designs for Food Research
Author(s) -
Hosahalli S. Ramaswamy,
C. Chen,
Sri Sreekanth,
Shyam S. Sablani,
Shiv O. Prasher
Publication year - 2000
Publication title -
journal of agricultural and marine sciences
Language(s) - English
Resource type - Journals
eISSN - 2410-1079
pISSN - 2410-1060
DOI - 10.24200/jams.vol5iss2pp97-106
Subject(s) - fractional factorial design , factorial experiment , factorial , artificial neural network , box–behnken design , design of experiments , mathematics , experimental data , statistics , computer science , artificial intelligence , response surface methodology , mathematical analysis
The ability of artificial neural networks (ANN) in predicting full factorial data from the fractional data corresponding to some of the commonly used experimental designs is explored in this paper. Factorial and fractional factorial designs such as L8, L9, L18, and Box and Behnken schemes were considered both in their original form and with some variations (L8+6, L15 and L9+1). Full factorial (3 factors x 5 levels) and fractional data were generated employing sixteen different mathematical equations (four in each category: linear, with and without interactions, and non-linear, with and without interactions). Different ANN models were trained and the best model was chosen for each equation based on their ability to predict the fractional data. The best experimental design was then chosen based on their ability to simulate the full- factorial data for each equation. In several cases, the mean relative errors with the L18 design (which had more input data than other models) were even higher than with other smaller fractional design. In general, the ANN assisted Lm, Box and Behnken, L15 and L18 designs were found to predict the full factorial data reasonably well with errors less than 5 %. The L8+6 model performed well with several experimental datasets reported in the literature.