Following the Perturbed Leader for Online Structured Learning

Authors: Alon Cohen, Tamir Hazan

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We complete our investigation with an online shortest path experiment and empirically show that our algorithm is both statistically and computationally efficient.
Researcher Affiliation Academia Alon Cohen ALONCOH@CSWEB.HAIFA.AC.IL University of Haifa, University of Haifa, Dept. of Computer Science, 31905 Haifa, Israel Tamir Hazan TAMIR@CS.HAIFA.AC.IL University of Haifa, University of Haifa, Dept. of Computer Science, 31905 Haifa, Israel
Pseudocode Yes Algorithm 1 Follow the perturbed leader Input: η > 0, set X {0, 1}d Set Θ1 0 for t = 1 to T do Sample γt N(0, I) Predict xt arg minx X x, Θt + ηγt Suffer loss xt, θt and accumulate Θt+1 Θt + θt end for
Open Source Code No The paper does not provide any links to or explicit statements about the release of source code for the methodology described.
Open Datasets No The paper states 'We constructed a directed graph with 53 vertices and 201 edges.' and describes how the adversary was built, but it does not refer to or provide access information for a publicly available or open dataset.
Dataset Splits No The paper describes an online learning setting and mentions evaluating algorithms with varying T, but it does not specify train/validation/test dataset splits. The experimental setup involves an online process rather than static data splits.
Hardware Specification No The paper does not specify any hardware details (e.g., GPU models, CPU types, memory) used for running the experiments.
Software Dependencies No The paper does not provide specific version numbers for any software components or libraries used.
Experiment Setup Yes The three algorithms we compare require to tune a parameter η. The minimax optimal choice of η depends on k, which in this case is the length of the longest s t path which is NP-hard and even NP-hard to approximate (Bj orklund et al., 2004). Therefore, we resort to choosing η based on the empirical performance of each of the algorithms against our adversary. We choose a nominal η by evaluating each algorithm on our problem with T = 100 on values of η ranging from en where n an integer is between 0 and 10 (we choose η from e n for Koolen et al. s algorithm). We run our experiments with horizons T between 1 and 100. We choose the nominal η and multiply it by p T/100 (divide by p T/100 for Koolen et al. s algorithm).