The Hedge Algorithm on a Continuum

Authors: Walid Krichene, Maximilian Balandat, Claire Tomlin, Alexandre Bayen

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

Reproducibility Variable Result LLM Response
Research Type Experimental We test our algorithm on a numerical example in R2 with convex quadratic loss functions of the form ℓ(t)(s) = 1 2(s µt) Qt(s µt) + ct restricted to the domain S R2 shown in Figure 2. ... For a simulation horizon of T = 104, we randomly generated the parameters µt, Qt and ct of the loss functions subject to the uniform bounds M = 10 and L = 5. We simulated the algorithm 2500 times for each of the different choices of the learning rates. Figure 3 shows means (solid lines), regret bounds (dashed lines) and regions between the 10% and 90% quantiles (shaded) of the per-round cumulative regret over these simulations.
Researcher Affiliation Academia Walid Krichene WALID@EECS.BERKELEY.EDU University of California, 652 Sutardja Dai Hall, Berkeley, CA 94720 USA Maximilian Balandat BALANDAT@EECS.BERKELEY.EDU University of California, 736 Sutardja Dai Hall, Berkeley, CA 94720 USA Claire Tomlin TOMLIN@EECS.BERKELEY.EDU University of California, 721 Sutardja Dai Hall, Berkeley, CA 94720 USA Alexandre Bayen BAYEN@BERKELEY.EDU University of California, 642 Sutardja Dai Hall, Berkeley, CA 94720 USA
Pseudocode Yes Algorithm 1 Hedge algorithm with initial density x(0) and learning rates (ηt). ... Algorithm 2 Dual averaging method with input sequence (ℓ(t)) and learning rates (ηt)
Open Source Code No The paper does not provide any explicit links or statements regarding the availability of open-source code.
Open Datasets No The paper uses randomly generated parameters and simulation for its numerical example, not a publicly available dataset. There is no mention of accessing a public dataset for training.
Dataset Splits No The paper describes a simulation-based numerical example with randomly generated parameters and a simulation horizon. It does not mention train/validation/test dataset splits.
Hardware Specification No The paper describes numerical simulations but does not specify any hardware (CPU, GPU, etc.) used to run these simulations.
Software Dependencies No The paper describes the algorithms and their theoretical properties and presents numerical simulation results, but it does not list any specific software dependencies with version numbers.
Experiment Setup Yes For a simulation horizon of T = 104, we randomly generated the parameters µt, Qt and ct of the loss functions subject to the uniform bounds M = 10 and L = 5. We simulated the algorithm 2500 times for each of the different choices of the learning rates.