This Too Shall Pass: Removing Stale Observations in Dynamic Bayesian Optimization

Authors: Anthony Bardou, Patrick Thiran, Giovanni Ranieri

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

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments establish the superiority of W-DBO, which outperforms state-of-the-art methods by a comfortable margin.
Researcher Affiliation Academia Anthony Bardou IC, EPFL Lausanne, Switzerland anthony.bardou@epfl.ch Patrick Thiran IC, EPFL Lausanne, Switzerland patrick.thiran@epfl.ch Giovanni Ranieri IC, EPFL Lausanne, Switzerland giovanni.ranieri@epfl.ch
Pseudocode Yes Algorithm 1 describes W-DBO, a DBO algorithm exploiting (12) to remove irrelevant observations on the fly.
Open Source Code Yes Its Python documented implementation can be found at https://github.com/WDBO-ALGORITHM/wdbo_ algo. A Py PI package can be quickly installed with the command pip install wdbo-algo.
Open Datasets Yes This benchmark comes from the temperature dataset collected from 46 sensors deployed at Intel Research Berkeley. It is a famous benchmark, used in other works such as [21, 22].
Dataset Splits No The paper describes initial observations and experimental setups but does not explicitly provide details about dataset splits for training, validation, or testing.
Hardware Specification Yes All experiments have been independently replicated 10 times on a laptop equipped with an Intel Core i9-9980HK @ 2.40 GHz with 8 cores (16 threads).
Software Dependencies No The paper mentions 'BOTorch [32] (MIT License)' and 'Py Bind11 [33] (BSD License)' but does not specify their version numbers, which are required for a reproducible description of software dependencies.
Experiment Setup Yes Each DBO algorithm begins its optimization task with 15 initial observations, uniformly sampled in S 0, 1/40 . At each iteration (at time t), (i) the noise level as well as the kernel parameters are estimated, and (ii) the GP-UCB acquisition function is optimized to get the next query. Unless stated otherwise, each DBO algorithm exploits a Matern-5/2 kernel as its spatial covariance function.