Learning unknown ODE models with Gaussian processes
Authors: Markus Heinonen, Cagatay Yildiz, Henrik Mannerström, Jukka Intosalmi, Harri Lähdesmäki
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the model s capabilities to infer dynamics from sparse data and to simulate the system forward into future. (Abstract)We employ 25 data points from one cycle of noisy observation data from VDP and FHN models, and 25 data points from 1.7 cycles from the LV model with a noise variance of σ2 n = 0.12. We learn the np ODE model with five training sequences using M = 62 inducing locations on a fixed grid, and forecast between 4 and 8 future cycles starting from true initial state x0 at time 0. (Section 4)We evaluate the method with two types of experiments: imputing missing values and forecasting future cycles. (Section 5) |
| Researcher Affiliation | Academia | 1Aalto University, Finland 2Helsinki Institute of Information Technology HIIT, Finland. |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The implementation is publicly available in http://www.github.com/cagatayyildiz/npode |
| Open Datasets | Yes | We use a benchmark dataset of human motion capture data from the Carnegie Mellon University motion capture (CMU mocap) database. The data used in this project was obtained from mocap.cs.cmu.edu. The database was created with funding from NSF EIA-0196217. |
| Dataset Splits | Yes | We perform model selection for lengthscales ℓby crossvalidation split of 80/20. |
| Hardware Specification | No | The paper does not specify any hardware details such as GPU models, CPU types, or memory used for experiments. It only mentions execution times: 'Training takes approximately 100 seconds per oscillator.' (Section 4) and 'The whole inference takes approximately few minutes per trajectory.' (Section 5). |
| Software Dependencies | No | The paper mentions 'We use the CVODES solver from the SUNDIALS package (Hindmarsh et al., 2005)' and 'We use an L-BFGS optimizer in Matlab.' However, specific version numbers for SUNDIALS or Matlab are not provided. |
| Experiment Setup | Yes | We employ 25 data points from one cycle of noisy observation data from VDP and FHN models, and 25 data points from 1.7 cycles from the LV model with a noise variance of σ2 n = 0.12. (Section 4) We learn the np ODE model with five training sequences using M = 62 inducing locations on a fixed grid... (Section 4) We place M = 53 inducing vectors on a fixed grid, and optimize our model starting from 100 different initial values... We use an L-BFGS optimizer in Matlab. (Section 5) We perform model selection for lengthscales ℓby crossvalidation split of 80/20. (Section 5) |