Influence of Learning Rate and Neuron Number on Prediction of Animal Phenotype Value Using Back-Propagation Artificial Neural Network

Abstract

In the past, a prediction equation based on the single nucleotide polymorphisms (SNP) is derived to calculate genomic breeding values (GEBV). However, the genome is very complex; a function could not reflect the relation between markers and phenotypes. Unlike the methods of regression, artificial neural networks (ANNs) could perform well for optimization in complex non-linear systems, however, artificial neural networks (ANNs) have not been used to calculate genomic breeding values (CEBV).In this paper, back-propagation neural network is used to predict the genomic breeding values (GEBV) or polygenic genotype value, and the different learning rate and hidden neurons number were used to discuss the influencing of the learning rate on estimating the polygenic genotype value. The result showed artificial neural networks could gather knowledge by detecting the relations between molecular marker polymorphism and phenotype value, and could predict the animal polygenic genotype value or breeding values as well as the molecular marker genotype being given. Training speed, prediction accuracy and stability could be improved along with enlargement of number of hidden neurons. The learning rate could not affect the prediction accuracy, and could almost affect the training speed. The training process was quite sensitive to the number of hidden neurons, even a hidden neurons change could lead to conspicuously training time prolong. It was necessary to have an applicable number of hidden neurons for predicting polygenic genotype value.