High Dimensional Bayesian Optimisation and Bandits via Additive Models

Authors: Kirthevasan Kandasamy, Jeff Schneider, Barnabas Poczos

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Via synthetic examples, a scientific simulation and a face detection problem we demonstrate that our method outperforms naive BO on additive functions and on several examples where the function is not additive.
Researcher Affiliation Academia Kirthevasan Kandasamy KANDASAMY@CS.CMU.EDU Jeff Schneider SCHNEIDE@CS.CMU.EDU Barnab as P oczos BAPOCZOS@CS.CMU.EDU Carnegie Mellon University, Pittsburgh, PA, USA
Pseudocode Yes Algorithm 1 Add-GP-UCB
Open Source Code Yes A Matlab implementation of our methods is available online at github.com/kirthevasank/add-gp-ucb.
Open Datasets Yes Here, we use galaxy data from the Sloan Digital Sky Survey to find the maximum likelihood values for 20 cosmological parameters. The likelihood is computed via an astrophysical simulation. Software is obtained from Tegmark et al (2006). ... In this experiment we use the VJ face dataset and the Open CV implementation (Bradski & Kaehler, 2008) which implements the classifier as a 22-stage cascade.
Dataset Splits No The paper initializes with Ninit points but does not specify explicit train/validation/test splits with percentages or sample counts for the evaluation datasets. The optimization process is sequential, where the "validation" is effectively part of the continuous evaluation of the acquisition function.
Hardware Specification No The paper does not provide specific hardware details such as CPU or GPU models used for running the experiments.
Software Dependencies No The paper mentions "Matlab implementation" and the "Di Rect" algorithm for maximizing the acquisition function, and "Open CV implementation" for face detection, but does not specify version numbers for these software components.
Experiment Setup Yes In our experiments, we set βt = 0.2 d log(2t) which offered a good tradeoff between exploration and exploitation. ... Following the recommendations in Bull (2011) we initialise Add-GP-UCB (and GP-UCB) using Ninit points selected uniformly at random. ... In our experiments, we choose the hyperparameters of the kernel by maximising the GP marginal likelihood (Rasmussen & Williams, 2006) every Ncyc iterations.