breedgym

BreedGym is an innovative reinforcement learning (RL) environment designed to optimize plant breeding programs, aiming to address the urgent global need for increased food production. Leveraging RL’s strengths in sequential decision-making, BreedGym simulates breeding program processes, enabling the design and evaluation of strategies to improve crop yield. The environment allows agents to interact with a custom-built breeding simulator developed using JAX, which offers high computational efficiency and adaptability. This simulator models the sequential processes of selection and mating based on genotype or phenotype states, with the goal of maximizing yield over multiple generations. BreedGym integrates RL and genetic optimization techniques, advancing research on AI-driven solutions for food security.

Initial results demonstrate BreedGym’s potential for transformative impacts. Compared to AlphaSimR, a widely-used breeding simulator, BreedGym achieves approximately 20x faster performance on a single GPU, critical for RL’s iterative training demands. By training a neural network to optimize a selection index, BreedGym has achieved a 7% improvement in yield estimation compared to traditional methods. These advancements represent the first AI-driven selection indices to outperform conventional approaches by a significant margin. Future directions include applying BreedGym to more realistic breeding schemas, developing adaptive strategies, and fostering collaboration between AI and agricultural sciences. BreedGym’s interdisciplinary approach offers a promising path forward in addressing the challenges of feeding a growing global population.