Deep Statistical Solvers
Authors: Balthazar Donon, Zhengying Liu, Wenzhuo LIU, Isabelle Guyon, Antoine Marot, Marc Schoenauer
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The third contribution is an experimental validation of the approach. The outline of the paper is the following. Section 5 experimentally validates the DSS approach, demonstrating its efficiency w.r.t. state-of-the-art solvers, and unveiling some super-generalization capabilities. |
| Researcher Affiliation | Collaboration | Balthazar Donon RTE R&D, INRIA, Université Paris-Saclay balthazar.donon@rte-france.com Wenzhuo Liu IRT System X wenzhuo.liu@irt-systemx.fr Antoine Marot RTE R&D antoine.marot@rte-france.com Zhengying Liu Université Paris-Saclay, INRIA zhengying.liu@inria.fr Isabelle Guyon Université Paris-Saclay, INRIA, Chalearn guyon@chalearn.org Marc Schoenauer INRIA, Université Paris-Saclay marc.schoenauer@inria.fr |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. It describes the architecture and process in text and diagrams. |
| Open Source Code | Yes | Our code is in the supplementary materials3, and links to the datasets are in references. 3code also available at https://github.com/bdonon/Deep Statistical Solvers |
| Open Datasets | Yes | The dataset [31] consists of 96180/32060/32060 training/validation/test examples from the distribution generated from the discretization of the Poisson equation... [31] Authors. Dataset of linear systems built from the poisson equation. http://doi.org/10. 5281/zenodo.4024811, 2020. |
| Dataset Splits | Yes | The dataset [31] consists of 96180/32060/32060 training/validation/test examples from the distribution generated from the Poisson equation... All free hyperparameters2 are tuned by trial and errors using the validation set, and all results presented are results on the test set. |
| Hardware Specification | Yes | Training is performed using the Adam optimizer [28] with the standard hyperparameters of Tensor Flow 1.14 [29], running on an Nvidia Ge Force RTX 2080 Ti. These algorithms are run on an Intel Xeon Silver 4108 CPU (1.80GHz)... State-of-the-art AC power flow computation uses the Newton-Raphson method, used as baseline here ([38] implementation, on an Intel i5 dual-core (2.3GHz)). |
| Software Dependencies | Yes | Training is performed using the Adam optimizer [28] with the standard hyperparameters of Tensor Flow 1.14 [29]... [38] implementation, on an Intel i5 dual-core (2.3GHz). |
| Experiment Setup | Yes | The number of updates k is set to 30 (average diameter size for the considered meshes). Each NN block has one hidden layer of dimension d = 10 and a leaky-Re LU non linearity; we have = 1e-3, lr = 1e-2 and γ = 0.9. The complete DSS has 49, 830 weights. Training is done for 280, 000 iterations (48h) with batch size 100. For case14 (resp. case118), k was set to 10 (resp. 30) ; we have = 1e-2, lr = 3e-3 and γ = 0.9 (resp. = 3e-4, lr = 3e-3 and γ = 0.9). The number of weights is 1, 722 for each of the k (M, D) blocks, hence 17, 220 (resp. 51, 660) in total. Training is done for 883, 000 (resp. 253000) iterations with batch size 1, 000 (resp. 500), and lasted 48h. |