Algorithm design for mixed-integer black-box optimization problems with uncertainty

Doctoral Thesis

The increasing competition in the automotive industry requires the tailored, swift development of technologically sophisticated vehicles. Therefore, the computationally expensive state-of-the-art simulation technologies are combined with optimization algorithms. An example of a real-world optimization problem is the task of designing parameters for vehicle dynamics control systems. This thesis investigates the design of optimization algorithms for solving complex and computationally expensive optimization problems. The thesis focuses on one central question: How can an optimization algorithm be effectively and efficiently designed to solve computationally expensive real-world problems? Therefore, a central aspect of the research is the development of a meta-optimization framework for tuning the parameters of the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). The proposed method involves the generation of surrogate optimization problems that mirror the complexity of the...Show moreThe increasing competition in the automotive industry requires the tailored, swift development of technologically sophisticated vehicles. Therefore, the computationally expensive state-of-the-art simulation technologies are combined with optimization algorithms. An example of a real-world optimization problem is the task of designing parameters for vehicle dynamics control systems. This thesis investigates the design of optimization algorithms for solving complex and computationally expensive optimization problems. The thesis focuses on one central question: How can an optimization algorithm be effectively and efficiently designed to solve computationally expensive real-world problems? Therefore, a central aspect of the research is the development of a meta-optimization framework for tuning the parameters of the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). The proposed method involves the generation of surrogate optimization problems that mirror the complexity of the original optimization problem landscapes. Then, the generated surrogate optimization problems are used as tuning references for the meta-optimization. Moreover, a dataset has been created that represents real-world optimization problem instances from the field of vehicle dynamics. This dataset reduces the computational effort associated with complex simulations and enables the efficient design of algorithms.
Show less

Parameter Tuning

CMA-ES

Vehicle Dynamics Design

Control Systems

Exploratory Landscape Analysis

Artificial Functions

Benchmarking