Abstract
This paper studies the regret bound of two transfer learning algorithms in Bayesian optimisation. The first algorithm models any difference between the source and target functions as a noise process, while the second explicitly models the bias between tasks. We provide the first formal theoretical proof that transfer learning speeds up Bayesian Optimisation by tightening regret bounds, demonstrating that prior knowledge effectively provides an "information head start."
Introduction
Bayesian Optimisation (BO) is a powerful framework for finding the global maximum of expensive-to-evaluate black-box functions. However, standard BO often starts from a "cold start," ignoring potentially valuable data from related previous experiments. While Transfer Learning (TL) has been applied empirically to address this, a rigorous theoretical understanding of how and why TL improves optimization efficiency has been lacking. This research bridges that gap by analyzing regret bounds—the cumulative loss incurred during optimization—within the Gaussian Process Upper Confidence Bound (GP-UCB) framework.
Methodology: Env-GP vs. Diff-GP
The authors analyze two distinct approaches to transferring knowledge from a source task to a target task:
- Env-GP (Envelope-Stretching): This method treats the discrepancy between source and target functions as stochastic noise. It "stretches" the confidence intervals of the source data until they are wide enough to encompass the target function. While robust, it is inherently pessimistic as it ignores any systematic structure in the task difference.
- Diff-GP (Difference-Modelling): This novel contribution explicitly models the bias function $g(x) = f_{target}(x) - f_{source}(x)$ as its own Gaussian Process. By learning the structure of the difference, Diff-GP can "correct" source data to better represent the target, offering a more optimistic and often more accurate model.
Theoretical Insights
The core contribution is the proof that both methods reduce the "maximum information gain" ($\gamma_T$) required to solve the target task. Effectively, source data provides a "head start," lowering the total information that must be extracted from the expensive target function.
Crucially, the authors establish a "safety net" theorem: if the source and target tasks are entirely unrelated (infinite shift), the regret bounds simply revert to the standard no-transfer case. This ensures that these TL methods do not perform worse than starting from scratch, mitigating the risk of "negative transfer."
Key Results
The algorithms were evaluated on synthetic 2D functions with varying "shift factors" and a real-world SVM hyperparameter tuning task using the UCI Image Segmentation dataset.
- Empirical Speedup: Both TL methods significantly outperformed standard BO across various task similarities.
- Diff-GP Superiority: As tasks became more distinct (moderate shifts), Diff-GP consistently outperformed Env-GP by successfully learning and compensating for systematic biases rather than just increasing uncertainty.
- Real-World Performance: In SVM tuning, Diff-GP achieved higher accuracy within a limited budget of 20 iterations in the majority of test cases.
Conclusion
This work provides a solid theoretical foundation for transfer learning in Bayesian Optimisation. By proving that TL leads to mathematically tighter regret bounds, the authors justify the use of prior data to accelerate expensive optimization tasks. Future work includes extending these proofs to other acquisition functions like Expected Improvement and developing automated methods for selecting the most relevant source tasks from large databases.
Summary updated