Vanilla Bayesian Optimization Performs Great in High Dimensions

Authors: Carl Hvarfner, Erik Orm Hellsten, Luigi Nardi

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate that standard BO works drastically better than previously thought for high-dimensional tasks, outclassing existing high-dimensional BO algorithms on a wide range of real-world problems. Further, we aim to shed light on the inner workings of the BO machinery and why minimal changes in assumptions yield a dramatic increase in performance. The result is a performant vanilla BO algorithm for dimensionalities well into the thousands. 6.1. Sparse Synthetic Test Functions We start by evaluating the DSP on a collection of commonly considered synthetic test functions with varying total and effective dimensionality. 6.3. High-dimensional Optimization Tasks We now benchmark Vanilla BO with the DSP against a collection of frequently considered tasks in the high-dimensional literature
Researcher Affiliation Collaboration 1Lund University, Lund, Sweden 2DBTune, Malm o, Sweden.
Pseudocode No The paper describes algorithms and methods in textual form and uses mathematical equations, but it does not include any clearly labeled pseudocode blocks or algorithm listings.
Open Source Code Yes Our code is publicly available at https://github.com/hvarfner/ vanilla_bo_in_highdim.
Open Datasets Yes We start by evaluating the DSP on a collection of commonly considered synthetic test functions with varying total and effective dimensionality... The Lunar Lander (12D) and Robot Pushing (14D) tasks from (Wang et al., 2017), as well as the Swimmer (16D) and Hopper (32D) reinforcement learning tasks from the Mu Jo Co suite (Todorov et al., 2012)... Specifically, we consider MOPTA08 (124D), SVM (388D), Lasso-DNA (180D), and the Mu Jo Co (Todorov et al., 2012) Ant (888D) and Humanoid (6392D) reinforcement learning tasks.
Dataset Splits No The paper mentions initialization (e.g., 'initialized with 30 samples') and acquisition optimization budgets, but it does not explicitly state how the datasets were split into training, validation, or test sets. While it uses standard benchmarks, the specific split percentages or methodology for their experiments are not detailed within the paper.
Hardware Specification No The paper mentions 'a single model fitting takes upwards of 5 minutes on SVM on 4 CPUS for later iterations.' However, it does not specify the model or type of these CPUs, nor does it mention any GPU models or other detailed hardware specifications.
Software Dependencies No The paper mentions software tools like 'Log EI (Ament et al., 2023)', 'Bo Torch (Balandat et al., 2020)', and 'GPy Opt-authors, 2016', but it does not provide specific version numbers for these or any other software dependencies, such as Python or specific libraries.
Experiment Setup Yes We instantiate the DSP with µ0 ? 3, which equates to ℓ 0.50 for D 6 under the mode of ppℓq. We initialize all methods with 30 samples, marked by a dashed vertical line. On all benchmarks, we use Log EI (Ament et al., 2023), using a low acquisition optimization budget of 512 initial (global) SOBOL samples and 512 Gaussian samples around the incumbent, followed by L-BFGS on the 4 best candidates, which is made possible by the low-complexity-high-smoothness model. As such, we fix σ2 f 1 to match the scale of the standardized observations, and to ensure that σ2 f does not diminish over time.