Year: 2026
Type: dissertation
Abstract: Offline model-based optimization (offline MBO) is a problem framework subsuming hard, yet essential optimization tasks from DNA sequence generation to robotic control and beyond. The key constraint is that only a fixed dataset of design-objective pairs is available in optimization. These problems can be framed as conditional generation, which motivates applying diffusion models, a powerful generative model family, to offline MBO. The goal of this thesis is twofold. First, it provides a literature survey of diffusion models and flow matching in single-objective offline model-based optimization. Second, it strives to illustrate the behavior of diffusion models through experiments from the lens of offline MBO. These experiments investigate varying guidance strength and model degradation under diminishing dataset quality. Results revealed that increasing the guidance strength leads to samples more concentrated in high-value regions of the design space. Interestingly, mean pairwise Euclidean distance, an evaluation metric for sample diversity, increased as well within low to moderate guidance scales suggesting a need for careful interpretation. Dataset experiments discovered that diffusion models appear to degrade gracefully as the dataset quality deteriorates. The learned model succeeded in generating samples guided towards desired regions, but the spread of samples increased as high-quality samples decreased in the training dataset. Finally, this work draws a connection between findings in dataset experiments and common trends in the literature, and identifies a demand for empirical guidelines of guided diffusion model performance.
Source: Aaltodoc (Aalto University)
Author: Eronen, Aapo
Institution: Aalto University
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7162229665