Bayesian Optimization for Personalized Vehicle Safety Design

Personalized system design aims to tailor engineering solutions to individual characteristics, enabling improved performance, safety, and user satisfaction across heterogeneous populations. This dissertation develops a Bayesian Optimization (BO) framework for simulation-based personalized design, with applications to enhance vehicle safety systems. Three major works have been explored. The first work develops the foundational framework for personalized policy optimization by proposing a functional, parameterized design policy that maps subject-specific covariates to optimal designs. This approach reduces the dimensionality of the searching space and enables tractable optimization across a heterogeneous population. A novel acquisition function, Personalized Information Gain (PIG), is developed and theoretically shown to restrict the search space to the trajectory of the optimal design policy, improving sampling efficiency. The second work addresses scalability and interpretability in the high-dimensional design spaces by introducing the Personalized Design Variable Score (PDVScore), which is a principled variable selection score that combines gradient-based sensitivity with population-level variability to identify important personalized design variables. This approach is supported by a theoretical justification linking the score to the Wasserstein-2 distance, offering a stable and interpretable measure of variable relevance. The third study extends the framework to a robust optimal solution that accounts for covariate measurement errors for practical deployment. A variance-based penalty term is incorporated into the optimization objective to reduce the performance gap between simulation and real-world conditions. This robust formulation is theoretically justified as an upper confidence bound on the deployed system response. These methodologies are evaluated through both computer simulations and a case study of designing a personalized high-fidelity vehicle restraint system. The results demonstrate substantial improvements in sampling efficiency, robustness to covariate uncertainty, and balanced safety performance across heterogeneous populations. Collectively, this work contributes a systemic and robust framework for scalable and reliable personalized system design under real-world deployment conditions.