Transfer Learning with Gaussian Processes for Bayesian Optimization

TL;DR

This paper explores hierarchical Gaussian process models for transfer learning in Bayesian optimization, introducing a new boosted GP transfer model and analyzing various methods' advantages and limitations through large-scale experiments.

Contribution

It provides a unified framework for hierarchical GP transfer models and introduces a novel boosted GP transfer model that balances complexity and performance.

Findings

The new boosted GP transfer model performs competitively with existing methods.

Hierarchical GP models offer a unified view and better understanding of transfer learning approaches.

Experimental results highlight the strengths and weaknesses of different transfer methods.

Abstract

Bayesian optimization is a powerful paradigm to optimize black-box functions based on scarce and noisy data. Its data efficiency can be further improved by transfer learning from related tasks. While recent transfer models meta-learn a prior based on large amount of data, in the low-data regime methods that exploit the closed-form posterior of Gaussian processes (GPs) have an advantage. In this setting, several analytically tractable transfer-model posteriors have been proposed, but the relative advantages of these methods are not well understood. In this paper, we provide a unified view on hierarchical GP models for transfer learning, which allows us to analyze the relationship between methods. As part of the analysis, we develop a novel closed-form boosted GP transfer model that fits between existing approaches in terms of complexity. We evaluate the performance of the different approaches in large-scale experiments and highlight strengths and weaknesses of the different transfer-learning methods.