Performance of Artificial Neural Networks (ANNs) and linear mixed models for prediction of breeding values

Document Type : Research Paper

Authors

• Samira Gavili ¹
• Mohammad Razmkabir ¹
• Amir Rashidi ¹
• Rezgar Arabzadeh ²

¹ Department of Animal Science, Faculty of Agriculture, University of Kurdistan, Sanandaj, Iran

² Department of Civil and Environmental Engineering, University of Waterloo, Waterloo, Ontario, Canada

Abstract

Keywords

• Artificial Neural Network
• Breeding Value
• Genetic Evaluation
• Markhoz Goat
• Weaning Weight

Main Subjects

• Animal Breeding and Genetics

Extended Abstract

Background and Objectives

Machine learning techniques, such as artificial neural networks (ANN) have been widely used for both prediction and classification tasks in many fields of knowledge; because they are quick, powerful, and flexible. ANNs are inspired by the human brain structure and function as if they are based on interconnected nodes in which simple processing operations take place. The spectrum of neural networks application is very wide, and it also includes agriculture. However, few studies are available on animal science (e.g., efficient farm management, classification of the quality of milk and meat products, discovery of mastitis, detection of pregnancy, animal diet formulation, verification of diseases, intelligent estrus control and prediction of 305 milk yields). The objective of this study was to prediction of breeding values of weaning weight in Markhoz goats based on the Mixed Model Equation (MME) and Artificial Neural Networks (ANNs).

Materials and Methods

Quality control and calculation of descriptive statistics was performed using the GLM procedure of the SAS statistical package. The pedigree file included 5541 kids produced by 261 bucks and 1616 does. CFC program was applied for pedigree monitoring and computation of inbreeding coefficients. In the first step, genetic evaluations and Best Linear Unbiased Prediction (BLUP) of breeding values for weaning weight was computed with the animal model by DMU program based on Mixed Model Equation (MME). Mixed model equations were derived by maximum likelihood from the joint normal probability distribution of the data and breeding values. Later, unique dataset was introduced to the ANN models by the R statistical program.

Results

A variety of models including, multilayer perceptron (MLP), radial basis function (RBF) and Support Vector Regression (SVR) were used to build the neural models. The artificial neural networks were trained and several networks (different hidden layers and nodes/ neurons) were evaluated. In artificial neural networks, the data were randomly divided to two parts (75% training and 25% for test/validation). Best architecture was selected according to the mean square of error and correlation. The final architecture was based on the two hidden layers for all models and the fitted numbers of neurons for MLP were 9 and 10, for RBF were 10 and 6, and for SVR were 7 and 10, respectively. The error and correlation factor are the two important performance metrics that helps to evaluate the performance of the ANN models.

Correlation between true breeding value (BVs predicted by MME) and estimated breeding value (BVs predicted by ANNs) for MLP, RBF and SVR models were 0.72, 0.49 and 0.73, respectively.

Conclusion

Analysis of farm data showed that the MLP and SVR models have higher performance than RBF for prediction of breeding values or ranking of individuals according to their genetic merit predictions. There is not sufficient evidence to support the application of ANNs for computation of EBV in animal breeding. More optimization of this novel methods, especially pedigree inclusion in analysis is recommended for future studies.