A0845
Title: Differentiable breeding: Automatic differentiation enables efficient gradient-based optimization of breeding strategies
Authors: Kosuke Hamazaki - Center for Advanced Intelligence Project (AIP), RIKEN (Japan) [presenting]
Hiroyoshi Iwata - The University of Tokyo (Japan)
Koji Tsuda - The University of Tokyo (Japan)
Abstract: Conventional breeding often requires extensive time to develop new cultivars, hindering rapid adaptation to global challenges. While genomic selection has accelerated breeding, there remains substantial room for improvement. Recent studies have explored complex decision-making in breeding schemes using various optimization techniques. However, these methods are challenged by constraints in simultaneously optimizing multiple parameters necessary for achieving more flexible optimization. To address these limitations, automatic differentiation of breeding schemes is implemented using PyTorch. By treating the entire breeding scheme as a differentiable computational graph, efficient gradient calculations are enabled for final genetic gains relative to progeny allocation parameters for each mating pair. Automatic differentiation is used to perform gradient-based optimization of progeny allocation strategies, aiming to maximize genetic gains in breeding schemes. The gradient-based strategy was then compared to black-box-based optimized and non-optimized strategies. The framework successfully outperformed the non-optimized strategy in terms of genetic gains, which demonstrates that the framework effectively harnessed gradient information via automatic differentiation.
