A consistently adaptive trust-region method

Authors: Fadi Hamad, Oliver Hinder

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

Reproducibility Variable Result LLM Response
Research Type Experimental We test our algorithm on learning linear dynamical systems [48], matrix completion [49], and the CUTEst test set [50]. Appendix D contains the complete set of results from our experiments.
Researcher Affiliation Academia Fadi Hamad Department of Industrial Engineering University of Pittsburgh Pittsburgh, PA 15261 fah33@pitt.edu Oliver Hinder Department of Industrial Engineering University of Pittsburgh Pittsburgh, PA 15261 ohinder@pitt.edu
Pseudocode Yes Algorithm 1: Consistently Adaptive Trust Region Method (CAT)
Open Source Code Yes Our method is implemented in an open-source Julia module available at https://github.com/ fadihamad94/CAT-Neur IPS.
Open Datasets Yes We test our algorithm on learning linear dynamical systems [48], matrix completion [49], and the CUTEst test set [50]. For our experiment, we use the public data set of Ausgrid, but we only use the data from a single substation. Details are provided in Appendix D.2.
Dataset Splits No The paper does not explicitly specify dataset splits (e.g., percentages or counts for training, validation, and test sets).
Hardware Specification Yes We perform our experiments using Julia 1.6 on a Linux virtual machine that has 8 CPUs and 16 GB RAM.
Software Dependencies No We perform our experiments using Julia 1.6 on a Linux virtual machine that has 8 CPUs and 16 GB RAM. (Only Julia version is given; specific versions for libraries like Optim.jl or CUTEst.jl are not provided.)
Experiment Setup Yes For these experiments, the selection of the parameters (unless otherwise specified) is as follow: r1 = 1.0, β = 0.1, θ = 0.1, ω = 8.0, γ1 = 0.0, γ2 = 0.8, and γ3 = 1.0. When implementing Algorithm 1 with some target tolerance ϵ, we immediately terminate when we observe a point xk with f(xk + dk) ϵ. This also includes the case when we check the inner termination criteria for the trust-region subproblem. The full details of the implementation are described in Appendix C. Our algorithm is stopped as soon f(xk + dk) is smaller than 10 5... We used 10000 as an iteration limit and any run exceeding this is considered a failure.