Resource limitations in optimizing multiple design parameters concurrently in time-intensive experiments can cause bottlenecks. Static resource-constrained optimization that determines the feasibility of the solution has received considerable attention in the literature. However, resources can restrict the optimization process itself, where the accessibility of the search space is skewed depending on the history of previous evaluations. When objective functions are expensive, Bayesian Optimization (BO) is a popular optimization technique with many advantages for real-world applications. The research in this thesis focuses on constraints in optimization processes where evaluations of the objective function have side effects on the accessibility of the search space, making different parts of it easier or harder to evaluate. We refer to these constraints as Dynamic Resource Constraints (DRCs) and note that, with some notable exceptions, they have received limited attention in the literature. First, we expand on the existing categorization of DRCs with additional types of constraints found in the literature and concrete industrial examples. In particular, we find that the structure of DRCs differs in severity, dependence on previous evaluations, and modeling in different fields such as mechanical engineering or drug design. Among the DRCs identified, we have decided to focus on the modular type, where some dimensions are harder to change than others, which have been largely overlooked in previous studies. Secondly, we empirically justify the advantages of BO, a popular sample-efficient algorithm for expensive problems, over some nature-inspired methods when dealing with this type of constraint, particularly in a modular setting on the tested problems. We observe that BO offers high adaptability for rapid changes in accessibility to the search space as it explores more across the constrained dimension than the other algorithms tested. Third, we develop and analyze various greedy and lookahead strategies for BO in dealing with DRCs. We find that BO greedy methods are still competitive to lookahead methods in the DRCs constrained modular setting. Finally, we propose a new problem type in the context of molecular discovery with DRCs. Our results suggest that resource-aware methods become less efficient compared to their resource ignorant algorithms as the problem size increases. Problems with DRCs face an exacerbated exploration-exploitation dilemma and a greater importance on long-term planning compared to standard constrained optimization.
- bayesian optimization
- expensive optimization
- resource constraints
- evolutionary algorithms
Tuning Bayesian Optimization for Dynamic Resource Constraints
Pricopie, S. (Author). 1 Aug 2025
Student thesis: Phd