Three Operator Splitting with a Nonconvex Loss Function

Authors: Alp Yurtsever, Varun Mangalick, Suvrit Sra

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we illustrate the effectiveness of the proposed method through numerical experiments on quadratic assignment problems. This section demonstrates the empirical performance of TOS on the quadratic assignment problem (QAP). Figure 1 compares the empirical performance of TOS and FW for solving (23) with chr12a and esc128 datasets from QAPLIB.
Researcher Affiliation Academia Alp Yurtsever 1 Varun Mangalick 1 Suvrit Sra 1 1Massachusetts Institute of Technology, Cambridge, Massachusetts, USA. Correspondence to: Alp Yurtsever <alpy@mit.edu>.
Pseudocode Yes Algorithm 1 Three Operator Splitting (TOS) Input: Initial point y1 Rn, step-size sequence {γt}T t=1 for t = 1, 2, . . . , T do zt = proxγtg(yt) xt = proxγth(2zt yt γt f(zt)) yt+1 = yt zt + xt end for Return: Draw τ uniformly at random from {1, 2, . . . , T} and output zτ.
Open Source Code Yes The source code is available online1. 1https://github.com/alpyurtsever/Nonconvex TOS
Open Datasets Yes Finally, we illustrate the effectiveness of the proposed method through numerical experiments on quadratic assignment problems. We evaluate TOS on the quadratic assignment problem using the well-known QAPLIB benchmark library (Burkard et al., 1997).
Dataset Splits No The paper evaluates the method on specific problem instances from the QAPLIB benchmark library, which are not typically split into training/validation/test sets in the manner of machine learning datasets. Therefore, it does not provide details on such dataset splits.
Hardware Specification Yes Experiments are performed in MATLAB R2018a on a Mac Book Pro Late 2013 with 2.6 GHz Quad-Core Intel Core i7 CPU and 16 GB 1600 MHz DDR3 memory.
Software Dependencies Yes Experiments are performed in MATLAB R2018a on a Mac Book Pro Late 2013 with 2.6 GHz Quad-Core Intel Core i7 CPU and 16 GB 1600 MHz DDR3 memory. For solving LAP, we employ an efficient implementation of the Hungarian method (Ciao, 2011).
Experiment Setup Yes For TOS (Algorithm 1), we output the last iterate instead of the random variable xτ. We use γt = 1/Lf step-size (Lf denotes the smoothness constant of f) instead of the more conservative step-size that our theory suggests (which depends on T). We start both methods from the same initial point y1. We construct y1 by projecting a random matrix with i.i.d. standard Gaussian entries onto the Birkhoff polytope via 1000 iterations of the alternating projections method.