Online Convex Optimization with Unbounded Memory

Authors: Raunak Kumar, Sarah Dean, Robert Kleinberg

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we present some simple simulation experiments.
Researcher Affiliation Academia Raunak Kumar Department of Computer Science Cornell University Ithaca, NY 14853 raunak@cs.cornell.edu Sarah Dean Department of Computer Science Cornell University Ithaca, NY 14853 sdean@cornell.edu Robert Kleinberg Department of Computer Science Cornell University Ithaca, NY 14853 rdk@cs.cornell.edu
Pseudocode Yes Algorithm 1: FTRL Input : Time horizon T, step size η, α-strongly-convex regularizer R : X R. ... Algorithm 2: Mini-Batch FTRL Input : Time horizon T, step size η, α-strongly-convex regularizer R : X R, batch size S.
Open Source Code Yes https://github.com/raunakkmr/oco-with-memory-code.
Open Datasets No We sample the disturbances {wt} from a standard normal distribution.
Dataset Splits No The paper describes simulation experiments where disturbances are sampled from a standard normal distribution. It does not mention traditional training, validation, or test dataset splits in the context of these simulations.
Hardware Specification No We run the experiments on a standard laptop.
Software Dependencies No We use the cvxpy library [Diamond and Boyd, 2016, Agrawal et al., 2018] for implementing Algorithm 1. The paper mentions the name of the library but does not specify its version number, nor does it list any other software dependencies with version numbers.
Experiment Setup Yes We set the time horizon T = 750 and dimension d = 2. We sample the disturbances {wt} from a standard normal distribution. We set the system matrix G to be the identity and the system matrix F to be a diagonal plus upper triangular matrix with the diagonal entries equal to ρ and the upper triangular entries equal to α. We run simulations with various values of ρ and α. ... We use step-sizes according to Theorems 3.1 and 3.3.