Trading Off Resource Budgets For Improved Regret Bounds

Authors: Thomas Orton, Damon Falck

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results: We benchmark both FPML and OGhybrid on an online black-box hyperparameter optimization problem based on the 2020 Neur IPS BBO challenge [Turner et al., 2021]. We find that both these new algorithms outperform OG for various compute budgets.
Researcher Affiliation Academia Damon Falck University of Oxford damon.falck@gmail.com Thomas Orton University of Oxford thomas.orton@cs.ox.ac.uk
Pseudocode Yes Algorithm 1 FPML(B,") Require: N B 1, " > 0. Initialize the cumulative cost C0(a) 0 for each arm a 2 A. for round t = 1, . . . , T do 1. For each arm a 2 A, draw a noise perturbation pt(a) 1 " Exp. 2. Calculate the perturbed cumulative costs for round t 1, Ct 1(a) Ct 1(a) pt(a). 3. Pull the B arms with the lowest perturbed cumulative costs according to Ct 1. Break ties arbitrarily. 4. Update the cumulative costs for each arm, Ct(a) Ct 1(a) + ct(a). end for
Open Source Code No The paper does not contain any explicit statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes the optimization problem at each round t 2 [T] is to choose the hyperparameters of either a multi-layer perceptron (MLP) or a lasso classifier for one of 184 classification tasks from the Pembroke Machine Learning Benchmark [Olson et al., 2017] (so T = 368).
Dataset Splits No No explicit train/validation/test dataset splits are provided for the tasks from the Pembroke Machine Learning Benchmark used in the experiments. The paper mentions running each algorithm setting 100 times, but this relates to statistical evaluation rather than data partitioning for model training/validation.
Hardware Specification No The paper mentions 'B available CPU cores' in the context of an application scenario but does not provide specific hardware details like GPU/CPU models, processor types, or memory amounts used for running its experiments.
Software Dependencies No The paper mentions using 'geometric sampling' and the 'Bayesmark package', but it does not specify version numbers for any software components or libraries.
Experiment Setup No The paper states that the " parameter for the bandit subroutines FPML-partial and Exp3 was set to their theoretically optimal values given B, N and T without any fine-tuning. However, it does not provide the specific numerical values of these parameters or other detailed hyperparameters for the experimental setup.