Safe Time-Varying Optimization based on Gaussian Processes with Spatio-Temporal Kernel

Authors: Jialin Li, Marta Zagorowska, Giulia De Pasquale, Alisa Rupenyan, John Lygeros

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show that TVSAFEOPT compares favorably against SAFEOPT on synthetic data, both regarding safety and optimality. Evaluation on a realistic case study with gas compressors confirms that TVSAFEOPT ensures safety when solving time-varying optimization problems with unknown reward and safety functions.
Researcher Affiliation Academia Jialin Li ETH Zürich (currently with UIUC) lijial@ethz.ch Marta Zagorowska NTNU (currently with TU Delft) m.a.zagorowska@tudelft.nl Giulia De Pasquale Eindhoven Univeristy of Technology g.de.pasquale@tue.nl Alisa Rupenyan Zürich University of Applied Sciences alisa.rupenyan@zhaw.ch John Lygeros ETH Zürich jlygeros@ethz.ch
Pseudocode Yes Algorithm 1 TVSAFEOPT
Open Source Code No The code accompanying the paper is currently under review and will appear shortly at https://www.research-collection.ethz.ch/.
Open Datasets Yes The data for the demand, compressor head, and degradation for the three compressors were obtained from [40] (Creative Commons Attribution Non Commercial Licence). [40] Marta Zagorowska. Degradation modelling in process control applications. Ph D thesis, Imperial College London, 2020. available at https://spiral.imperial.ac.uk/handle/10044/1/ 105173, online 22 May 2024.
Dataset Splits No The paper describes the datasets and problem setups but does not specify explicit training, validation, or test dataset splits in terms of percentages or counts, or by referencing predefined splits.
Hardware Specification Yes Experiments are conducted on an Intel i7-11370H CPU using Python 3.8.5.
Software Dependencies Yes The implementation utilizes the following libraries: GPy 1.12.0, NumPy 1.22.0, and Matplotlib 3.5.0.
Experiment Setup Yes The search space is X = [ 2, 2]2, uniformly quantized into 100 100 points. Both algorithms start with the singleton initial safe set {( 0.5, 0.0)}. The measurements are perturbed by i.i.d. Gaussian noise N(0, 0.012). The reward function is formulated as: f(x, t) = ex2 log(1 + y2) + 0.01t; The safety function is formulated as: c1(x, t) = 1 x + 0.5 0.5 1 cos 2π 6 2 y 0.3 0.5 1 cos 2π. ... σ1 1.0, σ2 = 25.0 for f, and σ2 = 15.0 for c1. SAFEOPT: ... σ3 1.0. ETSAFEOPT: ... sentivity of the event trigger δ as 0.01.