Development of an optimization framework for complex breeding programs based on the Modular Breeding Program Simulator (MoBPS)
by Azadeh Hassanpour
Date of Examination:2025-01-07
Date of issue:2025-05-16
Advisor:Prof. Dr. Henner Simianer
Referee:Prof. Dr. Stefan Scholten
Referee:Dr. Torsten Pook
Sponsor:This study was financially supported by BASF Belgium Coordination Center CommV.
Persistent Address:
http://dx.doi.org/10.53846/goediss-11274
Files in this item
Name:Thesis_AHassanpour.pdf
Size:2.45Mb
Format:PDF
Description:Thesis_AzadehHassanpour
The following license files are associated with this item:
Abstract
English
The aim of this thesis is to develop an optimization framework for complex breeding programs that involve numerous input parameters, auxiliary constraints, and stochastic outputs. Chapter 1 introduces the topic of stochastic simulation and optimization in the context of breeding program designs, and presents the key concepts essential to the subsequent chapters. In Chapter 2, kernel regression is introduced to manage stochasticity in evaluating objective functions through stochastic simulations. This approach addresses the challenge of optimizing breeding program designs using stochastic simulations, where each run represents only one of many possible outcomes, unlike deterministic methods with predictable results. Consequently, multiple replicates are necessary to enhance result precision. Breeding programs often involve numerous parameters, making it computationally expensive to simulate all possible designs multiple times. This limits optimization using stochastic simulations to a smaller set of scenarios. To address this limitation, a strategy employing kernel regression is proposed to efficiently explore a broader range of scenarios. The optimization process begins with identifying adjustable parameters for the breeding program and defining their feasible value ranges. A preliminary set of breeding designs is then randomly selected from all possible configurations, followed by initial simulations to gather baseline data and insights into potential solutions. Kernel regression is then applied to the data points representing different breeding program designs. This method fits a local regression curve, giving more weight to similar breeding schemes. It refines initial results by filtering noise and revealing underlying patterns, identifying potential optimal regions in the search space. By iteratively focusing on these regions, the search concentrates on the most effective breeding strategies, increasing the likelihood of finding the best design from thousands of possible scenarios. Although kernel regression has proven effective in optimizing breeding program designs and has been tested extensively across thousands of scenarios, these scenarios can differ in only a few parameters. As the number of parameters increases, the computational resources required for sufficient simulations across the entire search space grow substantially. Moreover, challenges emerged that required manual search space reduction through iterative steps and visual examination of potential optimal regions. To address these issues, Chapter 3 introduces a new evolutionary algorithm optimization framework to reduce the number of required simulations and operate fully automatically. The evolutionary algorithm begins similarly to kernel regression, with a random selection of breeding program designs and initial simulations. It then employs two main operators: selecting the best parameter settings based on their performance relative to the objective function and generating new breeding designs from these selected elite designs by combining two parameter settings or slightly modifying a single parameter setting. Kernel regression is integrated into the evolutionary algorithm to provide localized smoothing and mitigate stochastic effects. This algorithm effectively handles both continuous and class decision variables. The evolutionary algorithm is an iterative process that requires significant computational resources due to the need for multiple simultaneous simulations. To address this, the automation capabilities of the Snakemake workflow management system were integrated into the optimization approach, enabling parallel execution of simulations. The evolutionary algorithm was tested against the kernel regression method to evaluate its robustness. Results showed a 40-fold reduction in the number of simulations required compared to kernel regression, even when starting from a suboptimal search space or incorporating a binary decision variable. These findings highlight the practical advantages of using evolutionary algorithms for optimizing large-scale breeding program designs. Chapter 4 applies the optimization framework from Chapter 3 to enhance large-scale breeding programs with multiple parameters. The study focuses on two breeding programs modeled to reflect real-world scenarios, starting with a wheat line breeding program that includes six continuous variables, simulated using AlphaSimR. The second program is a complex hybrid wheat breeding program, simulated with MoBPS, involving 16 continuous parameters and one binary variable. Both programs demonstrated significant improvements in resource allocation within a fixed budget, showing that the framework enhances efficiency without requiring additional financial input or changing the core steps of the breeding process. A key strength of this study is that it is the first to apply the optimization framework to complex breeding programs with multiple parameters, including both continuous and class variables. Its application across different breeding programs (line and hybrid) using various simulators highlights the framework's adaptability to diverse breeding scenarios. This framework enables breeders to make informed, data-driven decisions before committing to costly and time-consuming real-world trials, allowing them to weigh alternatives and better predict outcomes. Chapter 5 summarizes the key highlights and main findings of this research. It discusses the potential applications of the developed algorithm in various aspects of breeding program design, as well as key considerations essential for optimizing breeding programs using stochastic simulations. Additionally, the chapter addresses potential challenges and limitations that may arise during the optimization process.
Keywords: optimization; stochastic simulation; breeding program design; evolutionary algorithm; kernel regression
