From EHVI to Diffusion Models: Pareto Set Learning for Expensive Multi-Objective Bayesian Optimization

Multi-objective Bayesian optimization (MOBO) is a framework for expensive black-box problems with multiple competing objectives, where each evaluation may require costly simulations or experiments. We begin with the basic geometry of Pareto fronts and the hypervolume indicator, then explain how Gaussian-process surrogate models and acquisition functions guide the search under a limited evaluation budget. The first half of the talk focuses on Expected Hypervolume Improvement (EHVI), one of the most widely used acquisition functions in MOBO.

The core of the talk is Expensive Multi-Objective Bayesian Optimization Based on Diffusion Models. We unpack CDM-PSL, a composite diffusion-model-based Pareto set learning method that combines unconditional and conditional sample generation, guided denoising, Gaussian-process surrogate modeling, and entropy-based objective weighting. We close with how diffusion models may complement classical EHVI-centered MOBO by generating diverse, high-quality candidate solutions in expensive evaluation regimes.

See Expensive Multi-Objective Bayesian Optimization Based on Diffusion Models for details.