Modelling the association between <i>in vitro</i> gas production and chemical composition of some lesser known tropical browse forages using artificial neural network

Abstract

In vitro gas production of four different browse plants (Azadirachta indica, Terminalia catappa, Mangifera indica and Vernonia amygdalina) was investigated under different extractions. The relationship between the forage composition parameters (dry matter, organic matter, crude protein, acid detergent fibre, neutral detergent fibre and acid detergent lignin), process parameters (extraction mode and incubation time), and volume of gas production were modelled with artificial neural network (ANN). The ANN model consisted of simple, multi-layered, back-propagation networks with eight input neurons consisting of the composition and process parameters and one output neuron for the gas volume. The networks were trained with different algorithms and varying number of layer and neuron in the hidden layer to determine the optimum network architecture. The network with single hidden layer having 45 ‘tangent sigmoid’ neurons trained with Livenberg-Marquard algorithm combined with ‘early stopping’ technique was found to be the optimum network for the model with R-value: mean = 0.9504; max. =
0.9618; min. = 0.9343; and std. = 0.0059. The influence of each chemical composition and processing parameters on gas production was simulated. The developed ANN model offers a more cost and time
efficient strategy in feed evaluation for ruminant animals.