Influence of Sample Size on Prediction of Animal Phenotype Value Using Back-Propagation Artificial Neural Network with Variable Hidden Neurons

Although linear multivariate approaches used to analyze large genetic data sets did not allow a large part of the total variance to be explained, strong distortions with nonlinear data sets, horseshoe effects had always been found. Artificial neural networks could gather their knowledge by detecting the patterns and relationships in data and learn through experience, and could perform well for optimization and prediction in complex non-linear systems. Artificial neural networks have been widely used in many life areas, but have not been used to predict the genomic breeding values or animal phenotypes. In this paper, Back-Propagation artificial neural network with Variable Hidden Neurons was used to predict the genomic breeding values. The results showed that artificial neural network could predict the animal genotype value, whatever there were interaction effect or not between gene loci. The sample size for training artificial neural network model could affect the training speed obviously, the training speed were obviously slowed along with enlargement of number of hidden neurons. A good structure of Back-Propagation artificial neural network needs a big sample for training its parameters. In some what, the sample size for training prediction model probably was not an important factor for prediction stability of artificial neural network; but large sample trained neural network model was very useful for training a Back-Propagation artificial neural network model with a small prediction error.