Physics-Informed Variational State-Space Gaussian Processes

Authors: Oliver Hamelijnck, Arno Solin, Theodoros Damoulas

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate our methods in a range of synthetic and real-world settings and outperform the current state-of-the-art in both predictive and computational performance.
Researcher Affiliation Academia Oliver Hamelijnck University of Warwick oliver.hamelijnck@warwick.ac.uk Arno Solin Aalto University arno.solin@aalto.fi Theodoros Damoulas University of Warwick t.damoulas@warwick.ac.uk
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Code to reproduce experiments is available at https://github.com/ohamelijnck/physs_gp.
Open Datasets Yes We follow [4] and use the dataset provided by D Asaro et al. [15] that has information from over 1, 000 buoys.
Dataset Splits No The paper mentions training and testing splits, but does not explicitly state a validation split for model selection or hyperparameter tuning. For example, for the Damped Pendulum: 'generate 20 points in t [0, 6] for training and 200 in t [6, 30] for testing'.
Hardware Specification Yes All models were run using an Nvidia Titan RTX GPU and an Intel Core i5 CPU.
Software Dependencies No The paper mentions software components like 'Adam' optimizer, but does not specify version numbers for any libraries, frameworks, or programming languages used in the implementation.
Experiment Setup Yes All were optimised for 1000 epochs using Adam [32] with a learning rate of 0.01. Both the GP and AUTOIP had an RBF kernel (following Long et al. [40]) and PHYSS-GP used a Matérn-7/2; all with a lengthscale of 1.0. The observation noise was initialised to 0.01 and the collocation 0.001. Both were fixed for the first 40% of training and then released.