Heuristic hyperparameter optimization of deep learning models for genomic prediction.

Area of Science:

• Quantitative genetics
• Animal breeding
• Machine learning

Background:

• Deep learning (DL) shows promise for genomic prediction in animal breeding.
• DL model performance is highly sensitive to hyperparameter settings.
• Current methods for hyperparameter optimization are often limited to discrete search spaces.

Purpose of the Study:

• To develop and evaluate an efficient method for optimizing hyperparameters in DL models for genomic prediction.
• To apply differential evolution (DE) for exploring complex hyperparameter spaces in DL models.
• To enhance the prediction accuracy and stability of genomic prediction models in livestock.

Main Methods:

• Utilized differential evolution (DE) for hyperparameter optimization in DL models.
• Applied the DE-optimized DL models to genomic prediction of livestock phenotypes using real genotype data.
• Evaluated model performance on pig and cattle datasets with simulated and real phenotypes.
• Compared DE-optimized models against those with "best practice" and randomly selected hyperparameters via cross-validation.

Main Results:

• DE-optimized DL models demonstrated superior predictive performance across all tested datasets.
• Optimized hyperparameters led to reduced overfitting in the DL models.
• The DE approach resulted in more consistent predictive performance across repeated model training.

Conclusions:

• Differential evolution is an effective strategy for optimizing DL hyperparameters in genomic prediction.
• This method enhances the accuracy and reliability of genomic prediction for livestock.
• Optimized DL models offer a significant advancement over current practices in quantitative genetics and animal breeding.