Parallel-in-Time Probabilistic Numerical ODE Solvers
Authors: Nathanael Bosch, Adrien Corenflos, Fatemeh Yaghoobi, Filip Tronarp, Philipp Hennig, Simo Särkkä
JMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Section 4, titled 'Experiments', describes the evaluation of the proposed parallel-in-time ODE filter on various ODEs, comparing its performance to both classic and probabilistic numerical ODE solvers. Figure 1, Figure 2, Figure 3, Figure 4, Figure 5 and Figure 6 all present empirical results, including RMSE, runtimes, and speed-up analyses. |
| Researcher Affiliation | Academia | All authors are affiliated with universities or public research institutions: 'Tübingen AI Center, University of Tübingen', 'Department of Electrical Engineering and Automation, Aalto University', and 'Center for Mathematical Sciences, Lund University'. All email addresses are academic domains. |
| Pseudocode | Yes | The paper includes 'Algorithm 1 Time-parallel Rauch Tung Striebel Smoother (Para RTS)' and 'Algorithm 2 Parallel-in-Time Probabilistic Numerical ODE Solver (Para IEKS)', which present structured pseudocode. |
| Open Source Code | Yes | Code for the implementation and experiments is publicly available on Git Hub.1 (Footnote 1: https://github.com/nathanaelbosch/parallel-in-time-ode-filters) |
| Open Datasets | Yes | The experiments are conducted on well-known ordinary differential equations (ODEs), which are mathematical models. The specific equations and initial conditions for these problems (Logistic ODE, Rigid Body Dynamics, Van der Pol Oscillator) are explicitly defined in Section 4 (Equations 42, 43, 44) within the paper, making the 'data' (the problem definitions) open and fully accessible. |
| Dataset Splits | No | The paper focuses on solving Ordinary Differential Equations (ODEs) numerically. These are mathematical problems defined by equations and initial conditions, not empirical datasets that are typically split into training, validation, and test sets. The concept of dataset splits is not applicable to the problem formulation or experimental setup in this paper. |
| Hardware Specification | Yes | Unless specified otherwise, experiments are run on an NVIDIA V100 GPU. The NVIDIA RTX 4090 used in this experiment has 16384 CUDA cores. |
| Software Dependencies | No | The paper mentions 'Python programming language', 'JAX software framework', 'Sci Py', and 'Diffrax' for implementation and reference solutions, but it does not specify version numbers for these software components, which is required for reproducibility. |
| Experiment Setup | Yes | Reference solutions are computed with diffrax s Kvaerno5 solver using adaptive steps and very low tolerances τ{abs,rel} = 10 12 (Section 4.1). In our experiments, we use a relative tolerance of 10 13 for the first criterion and absolute and relative tolerances of 10 9 and 10 6 for the second criterion, respectively (Section 3.2.2). |