Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
Bayesian optimization with derivatives acceleration
Authors: Guillaume Perrin, rodolphe le riche
TMLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This new acquisition criterion is found to improve Bayesian optimization on a test bed of functions made of Gaussian process trajectories in low dimension problems. The addition of first and second order derivative information is particularly useful for multimodal functions. Empirically investigating the effect of a new idea here adding derivatives acceleration on an optimization algorithm is difficult because the performance of an algorithm depends on both the function it is applied to and the tuning of its hyperparameters. The empirical tests we provide are designed to exclusively show the effects of the derivatives acceleration while avoiding all such experimental side effects. Section 4: Numerical experiments. |
| Researcher Affiliation | Academia | Guillaume Perrin guillaume.perrin@univ-eiffel.fr Université Gustave Eiffel, COSYS, 5 Bd Descartes 77454 Marne-La-Vallée, France. Rodolphe Leriche EMAIL LIMOS (CNRS, Mines Saint-Etienne, UCA) Saint-Etienne, France. |
| Pseudocode | Yes | Algorithm 1: Standard BO algorithm. Choose N0, budget, Y GPpµ, Cq ; Initialization... |
| Open Source Code | No | The paper mentions reviewing on OpenReview (openreview.net/forum?id=XXXX) and refers to a consortium (CIROQUO, https://doi.org/10.5281/zenodo.6581217) in the acknowledgments, but it does not provide an explicit statement or a specific repository link for the source code implementing the methodology described in this paper. |
| Open Datasets | No | The empirical tests we provide are designed to exclusively show the effects of the derivatives acceleration while avoiding all such experimental side effects. This is achieved firstly by testing on Gaussian processes whose hyper-parameters are known, therefore guaranteeing the compatibility of the model and the test function. We consider as test function class the set Fpdq θ of realizations of Zpdq θ that admits a global minimum strictly inside X (i.e., at a point of zero partial derivatives). |
| Dataset Splits | No | To begin, the function y is evaluated at N0 points uniformly chosen in X (typically according to a spacefilling design of experiments (Do E) Fang et al. (2006); Perrin & Cannamela (2017)). We note pxpnq, yn : ypxpnqqq N0 n 1 the obtained pairs. For each j P t1, 2u, and each k ě 2d 1, we generate 500 space-filling LHS made of k points in X Perrin & Cannamela (2017), which are written ! X pjq k,i )500 i 1. For each j P t1, 2u and each repetition of the experiment 1 ď i ď 500, the function yj D is evaluated at each point of r X pjq 3,i , and Algorithm 1 presented in Section 2 is run twice... |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used for running the experiments, such as GPU models, CPU types, or cloud computing specifications. |
| Software Dependencies | No | The paper mentions using the 'Nelder-Mead algorithm' for maximizing the acquisition criterion but does not specify any software libraries, frameworks, or their version numbers that were used for implementation or other parts of the experimental process. |
| Experiment Setup | Yes | The total number of calls to the objective function of each optimization run is equal to budget 100. The two types of searches are initialized with the evaluation of ypdq i,θ at the same space-filling LHS of dimension N0 3. The function µ is taken as a constant, and the function C is chosen in the class of tensorized Matérn kernels with smoothing parameter ν 5{2. The maximization of the acquisition criteria is performed in two steps: each acquisition criterion is first evaluated in 105 points randomly chosen in X, and 10 Nelder-Mead algorithms starting from the 10 most promising points among the random points are then launched in parallel to identify the new point at which to evaluate the objective function. |