Introduction
Optimization problems are ubiquitous in, e.g., natural sciences, IT & communications and banking. In a large portion of these problems the target function is a black-box, meaning that its internal structure is unknown, and expensive to evaluate. Bayesian optimization (BO) has emerged as an effective framework to address these problems: By employing a probabilistic machine learning model, Bayesian optimization estimates the underlying function with minimal assumptions regarding the functional form of the target. It leverages an acquisition function to guide the iterative optimization and minimize the number of target evaluations necessary.
Here, we propose two projects to extend the capabilities of the BOSS code [1,2] for BO developed by the AI-based Materials Science (AI4MS) group at TUM and the Materials Informatics Laboratory at Turku University.
Project 1: Noisy Acquisition Functions
Many applications involve noisy target functions that present a particular challenge for BO since conventional acquisition functions do not properly account for this noise, leading to suboptimal results. Research in this area is ongoing and while promising solutions have been proposed [3], widespread adoption is being hampered by a lack of testing and validation.
Project 2: Constrained optimization for materials
Real-world optimization problems typically feature constraints where certain resources, conditions, or regulations must be respected while still optimizing the target function. In particular for materials science, constraints on chemical composition and structure should be incorporated in predictive models to produce meaningful results. Despite theoretical developments in this area [4], applications to problems in materials science remain scarce.
Tasks:
(i) Review existing literature and identify promising solutions.
(ii) Implement suitable algorithms into the BOSS code.
(iii) Test and validate the implemented methods.
(iv) Apply the methods to data from ongoing research projects within the AI4MS group.
[1] https://sites.utu.fi/boss/
[2] M. Todorovic et al., npj Comput. Mater. 5, 35 (2019)
[3] B. Letham et al., Bayesian Anal. 14, 2, 495 - 519 (2019)
[4] J. Hernández-Lobato et al., J. Mach. Learn. Res. 17, 1, 5549–5601 (2016)
