back to list

Project: Breeding Program Optimization via Offline Reinforcement Learning

Description

Crop breeding programs aim to develop new cultivars with desirable traits through controlled mating within a population, enhancing agricultural productivity while reducing land use, greenhouse gas emissions, and water consumption. However, these programs face challenges like long turnover times, complex decision-making, long-term goals, and climate change adaptation.

Lower genotyping costs have generated extensive crop genomic data, presenting new opportunities for crop genetics adaptation. Advances in predicting phenotype traits from genomic data offer quicker and cheaper crop selection, avoiding the need for costly and slow trials. However, relying only on estimated traits risks creating low-diversity pools and compromising long-term success. Data-driven decision-making can help maintain genetic diversity and align with long-term goals such as climate adaptation.

In this project, we aim to model breeding programs using Offline Reinforcement Learning (RL) and develop methods leveraging crop genomic data to optimize breeding program design choices. Offline RL deals with the problems where simulation or online interaction is impractical, costly, and/or dangerous, allowing to automate a wide range of applications from healthcare and education to finance and robotics. However, learning new policies from offline data suffers from distributional shifts resulting in extrapolation error, which is infeasible to improve due to lack of additional exploration. Model-free RL algorithms that regularize the policies to stay close to the behavior policy, have limited generalization ability due to the sample complexity issue. Hence, model-based RL approaches that first learn the empirical MDP using the offline dataset and then freely explore in the learned environment for optimal policies, can effectively improve sample efficiency. 

Therefore, this project focuses on designing new or tailoring existing model-based offline RL techniques to optimize breeding programs. This involves conducting a comprehensive literature review to understand existing work in both domains and developing an approach to model the problem via reinforcement learning from offline data, learning policies that predict phenotype traits from genomic data for quicker and cheaper crop selection.

References:

• A Survey on Offline Reinforcement Learning: Taxonomy, Review, and Open Problems
• Breeding Programs Optimization with Reinforcement Learning
• MOPO: Model-based Offline Policy Optimization
• RAMBO-RL: Robust Adversarial Model-Based Offline Reinforcement Learning
• A Survey on Offline Model-Based Reinforcement Learning
  

Details

Supervisor

Maryam Tavakol

Secondary supervisor

IA

Ioannis Athanasiadis

External location

Wageningen University & Research

Interested?

Get in contact