Zeroth-Order Optimization for Composite Problems with Functional Constraints

Authors: Zichong Li, Pin-Yu Chen, Sijia Liu, Songtao Lu, Yangyang Xu7453-7461

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conduct numerical experiments to demonstrate the performance of our proposed ZO-i ALM. We consider the problem of resource allocation in sensor networks and the adversarial example generation problem.
Researcher Affiliation Collaboration Zichong Li1, Pin-Yu Chen2, Sijia Liu3, Songtao Lu2, Yangyang Xu1 1Department of Mathematical Sciences, Rensselaer Polytechnic Institute 2IBM Research 3Department of Computer Science and Engineering, Michigan State University
Pseudocode Yes Algorithm 1: Zeroth-order inexact augmented Lagrangian method (ZO-i ALM) and Algorithm 2: Zeroth-order accelerated proximal coordinate update for (15): ZO-APCU(G, H, µ, L, ε)
Open Source Code No The paper does not provide a direct link to source code or explicitly state that the code for the described methodology is publicly available.
Open Datasets Yes In the test, we use the ovarian cancer dataset (Conrads et al. 2004; Petricoin III et al. 2002) that are from m = 216 patients. Each data point has d = 4, 000 features and a label indicating whether the corresponding patient has ovarian cancer. We first use MATLAB s built-in lasso function (with λ = 0.01) to train a LASSO regression model parameterized by θ.
Dataset Splits No The paper describes the datasets used and some parameters but does not provide specific details on training, validation, or test dataset splits or splitting methodology.
Hardware Specification Yes All the tests were performed in MATLAB 2019b on a Macbook Pro with 4 cores and 16GB memory.
Software Dependencies Yes All the tests were performed in MATLAB 2019b on a Macbook Pro with 4 cores and 16GB memory.
Experiment Setup Yes We set d = 80, λ = 0.5, and ε = 0.5. ... In each call to the ZO-i PPM subroutine, we set the smoothness parameter to ˆLk = 50 + 0.3βk. We tune the parameters of ZO-Ada MM to α = 1, β1 = 0.75, β2 = 1, and fix the step size to 0.01 in ZO-Prox SGD. For each method, we choose a = 10 6 as the sampling radius and wk = 1/c(xk) as the dual step size. and In (23), we set λ = 0.01 and ε = 0.1. Due to the large variable dimension, we set ε = 1 in stopping conditions. ... In each method, we set a = 10 6 as the sampling radius and wk = 1/c(xk) as the dual step size.