EVALUATION OF NEURAL NETWORKS AS A TOOL FOR MANAGEMENT OF SWINE ENVIRONMENTS

Nine neural network configurations were developed and evaluated to pre dict the extent to which ambient temperature can be allowed to vary wi thout incurring excessive losses in rate of gain for ad-libitum-fed gr owing-finishing swine. The best network, chosen based on root mean squ are error, absolute error, and histograms of desired and target networ k outputs, was used in an experiment to determine maximum allowable am bient temperatures to achieve daily gain above 0.78 and 0.70 kg. Resul ts indicated that the network maintained constant growth rates (R2 gre ater-than-or-equal-to 0.99) of 0.79 and 0.78 kg/day compared to 0.93 k g/day under thermoneutral conditions, but the growth rate of animals i n the low growth rate treatment was considerably above the 0.70 kg/day target. Sensitivity analysis performed after the experiment showed th at the networks were not attempting to match daily gain goals. A neura l network, trained with a more comprehensive data set containing tempe rature increases and decreases, should improve upon the results found. Experimental results and sensitivity analysis of the simplest neural network developed also indicated a correlation among animal weight, gr owth limiting temperatures, and daily gain.