Decision support in breeding program in the Kenya Dairy cattle Prediction and Rule extraction using Neural Network System

Abstract

An important characteristic of all natural systems is the ability to acquire knowledge through experience and to adapt to new situations. This report presents findings on the use of neural networks to make predictions on the performance of daughter first lactation milk yield in Holstein Friesian cattle. Such prediction would ultimately lead to defining optimal breeding strategies. Data consisting of 6095 lactation records made by cows from 76 officially milk recorded Holstein Friesian herds and collected between 1988 and 2005 were used to predict the performance of the offspring based on some of the genetic traits of their parents using neural networks which was compared with linear regression model technique as baseline. The data was sorted using the current criteria for selection and a univariate animal model with relationships and the Derivative Free Restricted Maximum Likelihood procedure was used to predict the sire breeding values for milk yield. More accurate predictions were obtained by a neural network with 1 hidden unit than linear regression as can be seen by comparing the correlation coefficients and root mean square errors. This suggests that a non-linear relationship among the feature variables exists in the data and that these are learned by the hidden layer of the neural network. Feature sets which included sire information had high correlation coefficient. The black-box nature of neural networks was explained by extracting rules with domain expert and autoclass for both the continuous and the discrete valued inputs. Rules for discrete valued inputs for both categorization performed better on the 'low' and 'high' levels. This implied that performance at the two extremes are more important than average performance and that the client is particularly concerned with identifying mating with good potential and avoid matings with poor potential