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 and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Accelerating Model-Free Optimization via Averaging of Cost Samples
Authors: Guido Carnevale, Giuseppe Notarstefano
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical simulations on regression problems demonstrate that the proposed approach significantly outperforms existing model-free optimization methods. This section numerically validates our theoretical findings with Monte Carlo simulations of N = 20 trials in a logistic regression scenario (cf. Section 5.1) and a ridge regression one (cf. Section 5.2). |
| Researcher Affiliation | Academia | Guido Carnevale Department of Electrical, Electronic and Information Engineering Alma Mater Studiorum Universita di Bologna Bologna, Italy EMAIL Giuseppe Notarstefano Department of Electrical, Electronic and Information Engineering Alma Mater Studiorum Universita di Bologna Bologna, Italy EMAIL |
| Pseudocode | Yes | Algorithm 1 Averaged Model-Free Meta-Algorithm Initialization: x0 Rn, z1 0, . . . , z D 0 R for t = 0, 1, . . . do for j = 1, 2, . . . , D do if ℓ(xt + ϵdj) is gathered, i.e., sj t = 1 then zj t+1 = ℓ(xt + ϵdj) else zj t+1 = zj t end if end for xt+1 = xt γ PD j=1 gϵ(zj t+1, dj) end for |
| Open Source Code | No | We prefer to keep the code and data private for now, but we are open to sharing them upon request. |
| Open Datasets | No | In each trial, we randomly generate a synthetic dataset composed by m = 1000 points. See Appendix A.4 for details on the dataset generation. In this section, we describe how we generate the datasets used in the numerical simulations presented in Section 5. |
| Dataset Splits | No | The paper describes how synthetic datasets were generated but does not specify any training, test, or validation splits. It focuses on evaluating the algorithms on these generated datasets without explicit partitioning details. |
| Hardware Specification | No | The main body of the paper does not contain any specific hardware details used for running the experiments. |
| Software Dependencies | No | The main body of the paper does not provide specific software dependencies with version numbers. While the NeurIPS checklist mentions Matlab, it does not specify a version number, and this information is outside the main paper content. |
| Experiment Setup | Yes | We empirically select ϵ = 0.1 in all the schemes. We manually select the step sizes γ to achieve the fastest possible convergence for each method. Table 1 reports the chosen values. Finally, we set C = 1 and, in each trial, we randomly initialize the algorithms solution estimates by sampling x0 from a zero-mean Gaussian distribution with covariance matrix 10 In. We empirically select ϵ = 0.01 in both algorithms, while, as before, we select γ to achieve the fastest possible convergence of both methods, and report the chosen values in Table 1. |