Fast Bayesian Inference with Batch Bayesian Quadrature via Kernel Recombination

Authors: Masaki Adachi, Satoshi Hayakawa, Martin Jørgensen, Harald Oberhauser, Michael A Osborne

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we find that our approach significantly outperforms the sampling efficiency of both state-of-the-art BQ techniques and Nested Sampling in various real-world datasets, including lithium-ion battery analytics.
Researcher Affiliation Collaboration Masaki Adachi Machine Learning Research Group, University of Oxford Toyota Motor Corporation masaki@robots.ox.ac.uk Satoshi Hayakawa , Harald Oberhauser Mathematical Institute, University of Oxford {hayakawa,oberhauser}@maths.ox.ac.uk Martin Jørgensen, Michael A. Osborne Machine Learning Research Group, University of Oxford {martinj, mosb}@robots.ox.ac.uk
Pseudocode Yes Table 1 illustrates the pseudo-code for our algorithm. Rows 4–10 correspond to RCHQ, and rows 11–15 correspond to BQ. We can use the variance of the integral Var[Z|y] as a convergence criterion.
Open Source Code Yes Code: https://github.com/ma921/BASQ
Open Datasets Yes Hierarchical GP model The hierarchical GP model was designed for analysing the large-scale battery time-series dataset from solar off-grid system field data [3].
Dataset Splits No The paper describes evaluation metrics (KL, MAE, RMSE) and how batch sizes were optimized, but it does not specify traditional train/validation/test dataset splits for reproducibility.
Hardware Specification Yes 8Performed on Mac Book Pro 2019, 2.4 GHz 8-Core Intel Core i9, 64 GB 2667 MHz DDR4
Software Dependencies No The paper mentions the use of libraries like GPy, NumPy, SciPy, and Gurobi, and programming languages such as Python and MATLAB, but it does not provide specific version numbers for the software dependencies beyond the publication year of some referenced tools (e.g., Gurobi Optimizer Reference Manual, 2022).
Experiment Setup Yes Prior was set to a two-dimensional multivariate normal distribution, with a zero mean vector, and covariance whose diagonal elements are 2. The optimised batch sizes for each methods are BASQ: 100, batch WSABI: 16.