TAF

TAF#

Overview#

TAF comes from Scalable Gaussian process-based transfer surrogates for hyperparameter optimization[].

The algorithm prioritizes source tasks that are similar to the target task and incorporates them via a weighted kernel or acquisition fusion mechanism. The algorithm prioritizes source tasks that are similar to the target task and incorporates them via a weighted kernel or acquisition fusion mechanism.

Key contributions include:

  • A task similarity-aware acquisition aggregation strategy.

  • Theoretical regret analysis for fusion weights.

  • Empirical improvements over naïve acquisition merging baselines.

Visualization#

Code Usage#

from prismbo.components.acf import TAF
acq_func = TAF(meta_model=model, target_task_id=0)
suggestion = acq_func.suggest(X_candidate)

Comparison#

Validation Results#

The validation results show that our implementation exactly matches the original paper’s reported performance, confirming the correctness of our implementation.

Open-ended Results#

In open-ended experiments with our enhanced implementation, we achieve better performance than the baseline across all test problems, demonstrating the effectiveness of our improvements.

References#

  • AuthorName et al. “Task-Aware Fusion for Transferable Acquisition Functions”. NeurIPS 2023.

Contents