Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Breeding Programs Optimization with Reinforcement Learning

Omar G. Younis, Luca Corinzia, Ioannis N. Athanasiadis, Andreas Krause, Joachim M. Buhmann, Matteo Turchetta

TL;DR

The paper tackles optimizing crop breeding programs under climate-related pressures by reframing design decisions as sequential actions optimized with reinforcement learning. It introduces a Markov decision process formulation and a suite of Gym environments to simulate breeding, paired with an in silico maize dataset to benchmark methods. An RL agent trained with proximal policy optimization and curriculum learning achieves higher estimated genetic gain (e.g., ~6% SAM volume) than standard genomic selection in the simulated setting. The work provides practical tools and a path toward integrating RL into breeding, while noting the remaining gap between simulated and real-world outcomes and the need to manage large, dynamic decision spaces.

Abstract

Crop breeding is crucial in improving agricultural productivity while potentially decreasing land usage, greenhouse gas emissions, and water consumption. However, breeding programs are challenging due to long turnover times, high-dimensional decision spaces, long-term objectives, and the need to adapt to rapid climate change. This paper introduces the use of Reinforcement Learning (RL) to optimize simulated crop breeding programs. RL agents are trained to make optimal crop selection and cross-breeding decisions based on genetic information. To benchmark RL-based breeding algorithms, we introduce a suite of Gym environments. The study demonstrates the superiority of RL techniques over standard practices in terms of genetic gain when simulated in silico using real-world genomic maize data.

Breeding Programs Optimization with Reinforcement Learning

TL;DR

The paper tackles optimizing crop breeding programs under climate-related pressures by reframing design decisions as sequential actions optimized with reinforcement learning. It introduces a Markov decision process formulation and a suite of Gym environments to simulate breeding, paired with an in silico maize dataset to benchmark methods. An RL agent trained with proximal policy optimization and curriculum learning achieves higher estimated genetic gain (e.g., ~6% SAM volume) than standard genomic selection in the simulated setting. The work provides practical tools and a path toward integrating RL into breeding, while noting the remaining gap between simulated and real-world outcomes and the need to manage large, dynamic decision spaces.

Abstract

Crop breeding is crucial in improving agricultural productivity while potentially decreasing land usage, greenhouse gas emissions, and water consumption. However, breeding programs are challenging due to long turnover times, high-dimensional decision spaces, long-term objectives, and the need to adapt to rapid climate change. This paper introduces the use of Reinforcement Learning (RL) to optimize simulated crop breeding programs. RL agents are trained to make optimal crop selection and cross-breeding decisions based on genetic information. To benchmark RL-based breeding algorithms, we introduce a suite of Gym environments. The study demonstrates the superiority of RL techniques over standard practices in terms of genetic gain when simulated in silico using real-world genomic maize data.
Paper Structure (7 sections, 4 equations, 2 figures)

This paper contains 7 sections, 4 equations, 2 figures.

Table of Contents

  1. Introduction
  2. Problem definition
  3. Genetic simulation
  4. Gym environments
  5. Experiments
  6. Discussion
  7. Architecture details

Figures (2)

  • Figure 1: (On the left) Estimated SAM volume in $\mu m^3$ during the breeding program with standard and learned GS, averaged across 100 trials, with shaded area indicating standard error. (On the right) Estimated SAM volume percentage increase by applying the tricks explained in this section. Note: on "Curriculum learning" we also let the agent observe the generation number.
  • Figure 2: Representation of the policy net. The final score is a scalar value that is used by the environment to select the plants to cross.