Gaussian Process Bandit Optimisation with Multi-fidelity Evaluations

Authors: Kirthevasan Kandasamy, Gautam Dasarathy, Junier B. Oliva, Jeff Schneider, Barnabas Poczos

NeurIPS 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we demonstrate that MF-GP-UCB outperforms single fidelity methods on a series of synthetic examples, three hyper-parameter tuning tasks and one inference problem in Astrophysics.
Researcher Affiliation Academia Carnegie Mellon University, Rice University {kandasamy, joliva, schneide, bapoczos}@cs.cmu.edu, gautamd@rice.edu
Pseudocode Yes Algorithm 1 MF-GP-UCB Inputs: kernel κ, bounds {ζ(m)}M m=1, thresholds {γ(m)}M m=1.
Open Source Code Yes Our matlab implementation and experiments are available at github.com/kirthevasank/mf-gp-ucb.
Open Datasets Yes We used the regression method from [14] on the 4-dimensional coal power plant dataset. We tuned the 6 hyper-parameters the regularisation penalty, the kernel scale and the kernel bandwidth for each dimension each in the range (10 3, 104) using 5-fold cross validation. This experiment used M = 3 and 2000, 4000, 8000 points at each fidelity. ... We use Type Ia supernovae data [7] for maximum likelihood inference on 3 cosmological parameters
Dataset Splits Yes We set this up as a M = 2 fidelity experiment with the entire training set at the second fidelity and 500 points at the first. Each query was 5-fold cross validation on these training sets.
Hardware Specification No The paper mentions 'CPU Time' in its results (e.g., Figure 4) but does not provide specific details on the CPU, GPU, or any other hardware used for the experiments.
Software Dependencies No The paper mentions 'Our matlab implementation' but does not specify the version of Matlab or any other software dependencies with version numbers.
Experiment Setup Yes Our implementation uses some standard techniques in Bayesian optimisation to learn the kernel such as initialisation with random queries and periodic marginal likelihood maximisation. ... To set γ(m) s we use the following intuition: if the algorithm, is stuck at fidelity m for too long then γ(m) is probably too small. We start with small values for γ(m). If the algorithm does not query above the mth fidelity for more than λ(m+1)/λ(m) iterations, we double γ(m). ... The goal is to tune the kernel bandwidth and the soft margin coefficient in the ranges (10 3, 101) and (10 1, 105) respectively on a dataset of size 2000. ... We tuned the 6 hyper-parameters the regularisation penalty, the kernel scale and the kernel bandwidth for each dimension each in the range (10 3, 104) using 5-fold cross validation.