Interactive Hyperparameter Optimization in Multi-Objective Problems via Preference Learning

Authors: Joseph Giovanelli, Alexander Tornede, Tanja Tornede, Marius Lindauer

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In an experimental study targeting the environmental impact of ML, we demonstrate that our approach leads to substantially better Pareto fronts compared to optimizing based on a wrong indicator pre-selected by the user, and performs comparable in the case of an advanced user knowing which indicator to pick.
Researcher Affiliation Academia Joseph Giovanelli1, Alexander Tornede2, Tanja Tornede2, Marius Lindauer2 1Alma Mater Studiorum University of Bologna 2Institute of Artificial Intelligence, L3S Research Center, Leibniz University Hannover
Pseudocode No The paper does not contain a pseudocode block or a clearly labeled algorithm.
Open Source Code Yes All code including detailed documentation and the technical appendix can be found on Github1. 1https://github.com/automl/interactive-mo-ml
Open Datasets Yes The concrete hyperparameters our DNN algorithms exposes are defined by LCBench (Zimmer, Lindauer, and Hutter 2021), a well-known multi-fidelity deep learning benchmark, and given in Table 1 in the appendix. LCBench comprises evaluations of over 2000 funnel-shaped MLP neural networks with varying hyperparameters on 35 datasets of the Open ML CC-18 suite (Bischl et al. 2019).
Dataset Splits Yes The loss function can be used to assess the quality of a hyperparameter configuration by splitting the original dataset D into two disjoint datasets Dtrain and Dtest, where the model is trained only based on Dtrain but evaluated with L on Dtest. Overall, we seek to find the optimal hyperparameter configuration λ P Λ defined as λ P arg min λPΛ L p A p Dtrain, λq , Dtestq . (2) ... To this end, they split up a so-called validation dataset Dval from the training dataset Dtrain to estimate the loss Lp Ap Dtrain, λq, Dvalq and avoid a biased overfit to Dtest. ... We measure the accuracy of a model as the validation accuracy on 33% of the corresponding Open ML CC18 dataset.
Hardware Specification Yes The evaluations present in LCBench were performend on an Intel Xeon Gold 6242 with a maximum consumption of 150 Wh.
Software Dependencies No The paper mentions several software tools like SMAC, Optuna, Hyperopt, Hp Band Ster, Syne Tune, but does not specify their version numbers or other software dependencies with versions.
Experiment Setup Yes The concrete hyperparameters our DNN algorithms exposes are defined by LCBench (Zimmer, Lindauer, and Hutter 2021), a well-known multi-fidelity deep learning benchmark, and given in Table 1 in the appendix. ... We run both IB and PB for a budget of 30 evaluations on each of the datasets for 3 seeds and report the mean and standard deviation over the seeds and datasets.