Gespeichert in:
| 1. Verfasser: | |
|---|---|
| Format: | Recurso digital |
| Sprache: | |
| Veröffentlicht: |
Zenodo
2025
|
| Online-Zugang: | https://doi.org/10.5281/zenodo.17653598 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Inhaltsangabe:
- <h2>Abstract</h2> <p>Hyperparameter optimization (HPO) remains a critical bottleneck in machine learning pipelines, with existing methods lacking principled frameworks for understanding how hyperparameters affect model performance. We introduce Bifurcation-Based Hyperparameter Optimization (BBLA-HPO), building upon the foundational concepts of the Bifurcation-Based Learning Algorithm (BBLA) [ ], a novel approach that applies dynamical systems theory-specifically bifurcation theory-to guide hyperparameter search. This paradigm inversion treats hyperparameterperformance relationships as dynamical systems, where bifurcations represent critical transitions in model behavior. BBLA-HPO constructs a bifurcation tree over the hyperparameter space, using Upper Confidence Bound (UCB) selection and stability analysis to intelligently explore promising regions. Through comprehensive experiments on six diverse datasets spanning classification and regression tasks, we demonstrate that BBLA-HPO achieves competitive performance with random search while providing superior interpretability through bifurcation tree visualization and guaranteed ensemble diversity. On the California Housing regression task, BBLA-HPO achieves statistically significant improvements (+. %, p= .) over random search. Our results establish BBLA-HPO as a theoretically grounded, interpretable alternative to existing HPO methods, opening new research directions at the intersection of dynamical systems and AutoML.</p>