The power of first-order smooth optimization for black-box non-smooth problems

Authors: Alexander Gasnikov, Anton Novitskii, Vasilii Novitskii, Farshed Abdukhakimov, Dmitry Kamzolov, Aleksandr Beznosikov, Martin Takac, Pavel Dvurechensky, Bin Gu

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Experiments, 5.1. Reinforcement Learning, 5.2. Robust Linear Regression, 5.3. Support Vector Machine, Figure 1. Actor s reward for ADAM with Forward and Central differences for various γ and exact gradient ADAM. lr= 10 5. Figure 2. Loss for abalone scale dataset with batch size = 100, learning rate is 0.1 and γ = 10 5. Figure 3. Loss for a9a dataset with µ = 10 5, batch size = 100, lr = 0.1. γ = 10 5.
Researcher Affiliation Academia 1Moscow Institute of Physics and Technology, Dolgoprudny, Russia 2ISP RAS Research Center for Trusted Artificial Intelligence, Moscow, Russia 3Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE 4National Research University Higher School of Economics, Moscow, Russian Federation 5Weierstrass Institute for Applied Analysis and Stochastics, Berlin, Germany.
Pseudocode No The paper describes algorithms mathematically and in prose, but does not include any clearly labeled 'Algorithm' or 'Pseudocode' blocks.
Open Source Code Yes The code for all methods is available at https://github. com/OPTAMI/OPTAMI
Open Datasets Yes For the experiments we take the simple dataset "abalone scale" from the Lib SVM (Chang & Lin, 2011). We use the Lib SVM basic dataset "a9a" for our experiments with this problem.
Dataset Splits No The paper mentions training data for specific datasets (e.g., 'training is 3500 samples' for abalone scale) but does not provide explicit train/validation/test splits, percentages, or methodology for data partitioning across all experiments.
Hardware Specification Yes All the experiments were conducted in Python 3 and Py Torch 1.10.1 on an Ubuntu 20.04.3 LTS machine with Intel(R) Xeon(R) Silver 4215 CPU @ 2.50GHz and 125 GB RAM.
Software Dependencies Yes All the experiments were conducted in Python 3 and Py Torch 1.10.1 on an Ubuntu 20.04.3 LTS machine...
Experiment Setup Yes Figure 1. Actor s reward for ADAM with Forward and Central differences for various γ and exact gradient ADAM. lr= 10 5. Figure 2. Loss for abalone scale dataset with batch size = 100, learning rate is 0.1 and γ = 10 5. The Actor and Critic networks are made up of two hidden layers of fully-connected neural networks with h1 = 400 and h2 = 300 neurons with relu activation functions.