TECHNIQUES FOR DEVELOPMENT OF SWINE PERFORMANCE RESPONSE SURFACES

The NC204 swine growth model (NCPIG) was used to generate pig physiolo gical response data for a range of environmental variables during grow th (20 to 110 kg). These response data were then used to successfully train and validate three backward propagation neural network models de scribing the effect of environment on average daily gain, feed intake, heat production (total and fraction sensible), and physiological stat us of the animal. A generalization stage was conducted in which predic tions from NCPIG using actual weather data were compared to those foun d by the neural network models for the same environmental inputs. The neural network models were generally able to follow selected animal re sponse variables predicted from NCPIG, although average daily gain and daily feed intake exhibited occasional large deviations during the ge neralization phase, suggesting further training and validation are nee ded. The technique developed in this article shows how neural network models can be used to simplify data extraction from a complex model su ch as NCPIG by fitting neural networks to a few fundamental input rela tions based on carefully chosen numerical experiments. The simpler neu ral networks are then appropriate in instances where use of the full m odel is difficult or impossible, provided that parameters such as geno type and feed ration used for network training are maintained.