Dynamic Anisotropic Smoothing for Noisy Derivative-Free Optimization

Authors: Sam Reifenstein, Timothee Leleu, Yoshihisa Yamamoto

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the efficacy of our method through numerical experiments on artificial problems. Additionally, we show improved performance when tuning NP-hard combinatorial optimization solvers compared to existing state-of-the-art heuristic derivative-free and Bayesian optimization methods.
Researcher Affiliation Collaboration 1NTT Research Inc 2Stanford University. Correspondence to: Sam Reifenstein <Sam.Reifenstein@nttresearch.com>.
Pseudocode Yes Algorithm 2
Open Source Code Yes An implementation for our algorithm can be found in the following Git Hub repository: https: //github.com/Sam1234567/DAS-Autotuner? tab=readme-ov-file#das-autotuner
Open Datasets No The paper uses generated problem instances (e.g., 'random 3-SAT', 'Sherrington-Kirkpatrick (SK) model') based on defined parameters and models, rather than explicitly mentioning or linking to a pre-existing, publicly available or open dataset.
Dataset Splits No The paper does not explicitly describe train/validation/test dataset splits in the context of model training. It describes generating problem instances and evaluating performance over multiple realizations or trajectories, which is different from standard dataset splitting methodologies.
Hardware Specification No The paper mentions parallel computation 'such as in a GPU' but does not provide specific hardware details (e.g., CPU or GPU models, memory specifications) used for running the experiments.
Software Dependencies No No specific software dependencies with version numbers (e.g., programming languages, libraries, frameworks, or solvers) were mentioned in the paper.
Experiment Setup Yes The parameters for this algorithm are B0 (initial batch size), γ (batch size exponent) and t which is the integration time step as well as αL and αx in eq. (9). For all results in this paper, we use αL = 1/D, αx = 1 whereas the other parameters depend on the amount of parallelism that can be used and the properties of f. ... we always use wmax = 2, and, except for the benchmark results on the Rosenbrock function (section 3.2), we use wmin = 0. ... tuning the SAT solver with N = 150, α = 4.0 T = 148. ... N = 150, βE = 0.01 (left) and N = 300, βE = 0.005 (right).