Multi-task / transfer learning Bayesian optimization for heterogenous search spaces

[sgbaird]

There’s an interest from the community (e.g., BayBE) in being able to perform transfer learning within a Bayesian optimization framework to avoid the “cold-start” characteristic of many Bayesian optimization campaigns despite related data being available. Some examples:

1. Maximizing chemical reaction yields based on a related chemical reaction yield that shares a subset of parameters
2. Maximizing measured superhardness while incorporating physics-based simulations of bulk and shear moduli simulations (related parameters, search space between simulation and experiment is not one-to-one)
3. Maximizing experimentally observed yield strength and elongation at fracture while incorporating physics-based simulations (e.g., Thermo-Calc)

Some questions come up, such as:

1. What are effective ways to approach benchmarking with this topic?
2. What are the state-of-the-art (already implemented) methods for doing so?

Related resources:

Multi-task vs. multi-fidelity: which should I choose?

Basic transfer learning example · emdgroup/baybe · Discussion #257

General feedback - documentation, functionality, and software dev ·...

[sgbaird]

Some related literature based on Meta AE workshop 2024:

• Fan, Z.; Han, X.; Wang, Z. HyperBO+: Pre-Training a Universal Hierarchical Gaussian Process Prior for Bayesian Optimization.
• Fan, Z.; Han, X.; Wang, Z. Transfer Learning for Bayesian Optimization on Heterogeneous Search Spaces. arXiv February 13, 2024. [2309.16597] Transfer Learning for Bayesian Optimization on Heterogeneous Search Spaces.
• Astudillo, R.; Frazier, P. I. Bayesian Optimization of Function Networks. arXiv December 31, 2021. [2112.15311] Bayesian Optimization of Function Networks.
• Astudillo, R.; Frazier, P. Bayesian Optimization of Composite Functions. In Proceedings of the 36th International Conference on Machine Learning; PMLR, 2019; pp 354–363.

Additionally:

• Ax: Multi-task BO
• BoTorch: Composite BO
• Composite Bayesian Optimization with the High Order Gaussian Process
• Composite Multi-Task BO
• Meta-Learning with the Rank-Weighted GP Ensemble (RGPE)

Key terms:

• composite functions
• function networks
• multi-task
• meta-learning

[jakobzeitler]

We have done a lot of work in this area, see: MATTERHORN STUDIO - robust-TLBO IP

Keen to share more later this year as we embark on some case studies.

Happy to chat and discuss!