Hyperparameter Optimization via Interacting with Probabilistic Circuits

Jonas Seng, Fabrizio Ventola, Zhongjie Yu, Kristian Kersting

Keywords: hyperparameter optimization, model-based optimization, tractable probabilistic modeling

Abstract: Despite the growing interest in designing truly interactive hyperparameter optimization (HPO) methods, to date, only a few allow to include feedback from experts. However, these methods add friction to the interactive process, rigidly requiring to fully specify the expert input as prior distribution ex ante and often imposing additional constraints on the optimization framework. This hinders the flexible incorporation of expertise and valuable knowledge of domain experts, which might provide partial feedback at any time during optimization. To overcome these limitations, we introduce a novel Bayesian optimization approach leveraging tractable probabilistic models named probabilistic circuits (PCs) as surrogate model. PCs encode a tractable joint distribution over the hybrid hyperparameter space and enable exact conditional inference and sampling, allowing users to provide valuable insights interactively and generate configurations adhering to their feedback. We demonstrate the benefits of the resulting interactive HPO through an extensive empirical evaluation of diverse benchmarks, including the challenging setting of neural architecture search.

Submission Checklist: Yes

Broader Impact Statement: Yes

Paper Availability And License: Yes

Code Of Conduct: Yes

Optional Meta-Data For Green-AutoML: This blue field is just for structuring purposes and cannot be filled.

Evaluation Metrics: No

Submission Number: 12