Artificial neural networks approach for predicting methionine requirement in broiler chickens

Abstract

The objective of this research was to apply artificial neural networks (ANNs) for predicting the methionine
requirement in broiler chickens at day 1-10 (starter period) and day 11-21 (grower period). A total of 28 data was
obtained from five hundred and sixty male broiler chicks (Ross 308), which were divided into twenty-eight pens with
twenty chickens in each. Body weight was determined at days 10 and 21. A bootstrapping method was used to multiply
the observations to overcome the limited data for training. A total of 280 data was obtained and divided into a training
set (n = 220) and a testing set (n = 60). The level of TSAA supplementation (%) was used as a variable in the input node,
whereas the average daily gain (g) was used as a variable in the output node. The model evaluation was determined
by R2, mean absolute deviation (MAD), mean absolute percentage error (MAPE) and mean square error (MSE).
Quadratic regression and ANNs with radial basis function were used to develop the model using Python programing.
The results showed that no significant difference (P>0.05) was observed in means between the original data and the
bootstrapping data. The ANNs showed greater accuracy of R2 when compared with quadratic regression at the starter
(0.7178 vs. 0.7294) and grower (0.8086 vs. 0.8097) periods. For error measurements, ANNs also resulted in lower MAD,
MAPE and MSE when compared with quadratic regression at the starter and grower periods. In conclusion, the optimal
level of methionine (as total sulphur amino acids) obtained by ANNs was 1.13 and 0.99% for starter and grower periods,
respectively. Therefore, ANNs are an alternative method to predict methionine requirements of broiler chickens for
improving poultry production.