Multi-Objective Bayesian Optimization with Asynchronous Batch Selection

Abstract

Multi-objective optimization problems are widespread in scientific, engineering, and design f ields, necessitating a balance of trade-offs between conflicting objectives. These objectives often represent black-box functions, which are costly and time-consuming to evaluate. Multiobjective Bayesian optimization (MOBO) offers a valuable approach to guide the search for optimal solutions. To enhance efficiency, batch evaluations are employed to test multiple samples simultaneously, aiming to further reduce evaluation times. However, in scenarios involving varying evaluation times, standard batch strategies often lead to suboptimal resource utilization and inefficiencies. Asynchronous evaluations emerge as a promising solution to optimize resource usage under these conditions. Despite their potential, there has been no prior work or method specifically tailored to address asynchronous evaluations within the MOBO framework. To bridge this critical gap, this thesis proposes a comprehensive adaptation and analysis of existing Bayesian optimization methods for asynchronous MOBO scenarios. It also introduces a novel selection strategy, α-HVI, empirically validated through tests on both synthetic and real-world functions.