Symplectic Spectrum Gaussian Processes: Learning Hamiltonians from Noisy and Sparse Data
Authors: Yusuke Tanaka, Tomoharu Iwata, naonori ueda
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on several physical systems show that SSGP offers excellent performance in predicting dynamics that follow the energy conservation or dissipation law from noisy and sparse data. Data. We evaluated the proposed model, SSGP, using two physical systems: pendulum, and Duffing oscillator. |
| Researcher Affiliation | Industry | Yusuke Tanaka Tomoharu Iwata Naonori Ueda NTT Communication Science Laboratories |
| Pseudocode | No | The paper describes the inference procedure and other steps in prose, but it does not include any formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/yusuk-e/SSGP |
| Open Datasets | No | We generated trajectory data by employing a numerical integrator, i.e., the Dormand Prince method with adaptive time-stepping, implemented in torchdiffeq5 [3, 4]. |
| Dataset Splits | Yes | We randomly split the trajectory data and used 70% for training and 30% for validation. |
| Hardware Specification | Yes | The average training time when setting M = 250 was 2943.0 seconds for the dataset of the pendulum with friction; the experiments were conducted on the AMD EPYC 7313 CPU (3.0GHz). |
| Software Dependencies | No | The paper mentions "implemented in Py Torch [30]" and "torchdiffeq5 [3, 4]" but does not specify version numbers for these software dependencies, which would be necessary for full reproducibility. |
| Experiment Setup | Yes | We trained the model using the Adam optimizer [21] with learning rate of 10 3 for 104 epochs, implemented in Py Torch [30]. We performed numerical integration by the adaptive Dormand Prince method [3, 4] with the relative and absolute tolerances of 10 8. We set the numbers of Monte Carlo samples to K = 1 and L = 100. The number M of spectral points was chosen from {100, 250, 500} based on the validation error. |