Black-Box Optimization with Local Generative Surrogates
Authors: Sergey Shirobokov, Vladislav Belavin, Michael Kagan, Andrei Ustyuzhanin, Atilim Gunes Baydin
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In Section 4 we evaluate L-GSO on a set of toy problems and compare it with frequently used methods including numerical differentiation, Bayesian optimization, and score function-based approaches, and present results of a realistic use case in the high-energy physics domain. |
| Researcher Affiliation | Academia | Sergey Shirobokov Department of Physics Imperial College London United Kingdom s.shirobokov17@imperial.ac.uk Vladislav Belavin National Research University Higher School of Economics Moscow, Russia vbelavin@hse.ru Michael Kagan SLAC National Accelerator Laboratory Menlo Park, CA United States Andrey Ustyuzhanin National Research University Higher School of Economics Moscow, Russia Atılım Güne s Baydin Department of Computer Science Department of Engineering Science University of Oxford United Kingdom |
| Pseudocode | Yes | Algorithm 1 Local Generative Surrogate Optimization (L-GSO) procedure |
| Open Source Code | No | The paper mentions open-source tools used (e.g., PyTorch, Pyro, GEANT4, Fair Root) by citing them, but it does not provide an explicit statement or link to the authors' own implementation code for the described methodology. |
| Open Datasets | Yes | Neural Network Weights Optimization Problem In this problem, we optimize neural network weights for regressing the Boston house prices dataset [28]. |
| Dataset Splits | No | The paper mentions using several datasets/problems (e.g., 'Boston house prices dataset', 'Rosenbrock problem') and performing multiple runs, but it does not provide specific details on how these datasets were split into training, validation, and test sets. |
| Hardware Specification | No | The paper mentions that 'The reported study utilized the supercomputer resources of the National Research University Higher School of Economics' but provides no specific details such as CPU/GPU models, memory, or other hardware specifications. |
| Software Dependencies | No | The paper mentions using 'Py Torch [46] and Pyro [9]' for gradient calculations and 'GEANT4 [2] and Fair Root [3]' for physics simulation, but it does not provide specific version numbers for any of these software dependencies. |
| Experiment Setup | Yes | We tune the hyper-parameters of all baselines for their best performance. For all experiments we run L-GSO and baselines ten times with different random seeds and show the averages and standard deviations in Figures 2 and 4. We use the same GAN architecture with 60k parameters for all experiments (for model details see Appendix A). The value of ϵ = 0.2 was found to be robust and work well in all experiments except the 'three hump problem' which required a slightly larger value of ϵ = 0.5. |