Weak Form Generalized Hamiltonian Learning

Authors: Kevin Course, Trefor Evans, Prasanth Nair

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Numerical Studies We compare our approach (GHNN) to a fully connected neural network (FCNN) and Hamiltonian neural network (HNN). All models were trained on an Nvidia Ge Force GTX 980 Ti GPU. ... The error in the states, its derivatives, and training time for the three methods of parameter estimation are given in Table 1.
Researcher Affiliation Academia Kevin L. Course University of Toronto kevin.course@mail.utoronto.ca Trefor W. Evans University of Toronto trefor.evans@mail.utoronto.ca Prasanth B. Nair University of Toronto pbn@utias.utoronto.ca
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Code can be found online at: https://github.com/coursekevin/weakformghnn.
Open Datasets No The paper generates synthetic data by collecting 'measurements of the pendulum state corrupted by Gaussian noise' or 'measurements of the state corrupted by Gaussian noise' from well-known dynamical systems (nonlinear pendulum, Lorenz 63, Duffing oscillator) rather than using pre-existing public datasets with explicit access information.
Dataset Splits No The paper describes collecting measurements for training and evaluation but does not specify explicit train/validation/test dataset splits (e.g., percentages, sample counts, or predefined split references).
Hardware Specification Yes All models were trained on an Nvidia Ge Force GTX 980 Ti GPU.
Software Dependencies No The paper mentions 'Py Torch' and 'torchdiffeq' but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes Unless otherwise noted, we will use the default settings for the adjoint ODE solvers offered in Chen et al. s package; at the time of writing, this includes a relative tolerance of 10 6 and an absolute tolerance of 10 12 with a Runge-Kutta(4)5 adaptive ODE solver. ... For all experiments that use weak derivative regression, the test space is spanned by 200 evenly spaced Gaussian radial basis functions with a shape parameter of 10 over each mini-batch integration window;