Online Linear Quadratic Control

Authors: Alon Cohen, Avinatan Hasidim, Tomer Koren, Nevena Lazic, Yishay Mansour, Kunal Talwar

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 7. Experiments We demonstrate our approach on the problem of regulating conditions inside a data center (DC) server floor in the presence of time-varying power costs. We learn system dynamics from a real data center, but vary the costs and run algorithms in simulation. ... The normalized regret 1 T RT of to the two strategies is shown in Fig. 2.
Researcher Affiliation Collaboration 1Google Research, Tel Aviv 2Technion Israel Inst. of Technology 3Bar Ilan University 4Google Brain, Mountain View 5Tel Aviv University. Correspondence to: Alon Cohen <aloncohen@google.com>, Tomer Koren <tkoren@google.com>.
Pseudocode Yes Algorithm 1 Online LQ Controller
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets No We learn system dynamics from a real data center, but vary the costs and run algorithms in simulation. We learn a linear approximation (A, B) of the dynamics in the operating range of interest on 4h of exploratory data with controls following a random walk.
Dataset Splits No The paper does not provide specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology).
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes We learn system dynamics from a real data center, but vary the costs and run algorithms in simulation. ... We amplify the noise by a factor of 5. We set the diagonal coefficients of Qt corresponding to the most recent (normalized) sensor measurements to 1 and remaining coefficients to 0, and keep Qt = Q constant throughout the experiment. We set diagonal coefficients of Rt corresponding to water usage (valve command) to 1 throughout, and all coefficients corresponding to power usage (fan speed) to rt. We generate rt by (a) i.i.d sampling a uniform distribution on [0.1, 1], and (b) using a random walk restricted to [0.1, 1] taking steps of size 0.1, 0.1, 0 with probabilities 0.1, 0.1, 0.8 respectively. We run the FLL algorithm on this problem with the following modifications: we set Qp 1 = Q, and Rp 1 = Ik, an upper bound on Rt.