Neural Solvers for Fast and Accurate Numerical Optimal Control
Authors: Federico Berto, Stefano Massaroli, Michael Poli, Jinkyoo Park
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The performance is evaluated in direct and receding horizon optimal control tasks in both low and high dimensions, where the proposed approach shows consistent Pareto improvements in solution accuracy and control performance.In particular, we then carry out performance and generalization evaluations in direct and model predictive control tasks. |
| Researcher Affiliation | Academia | Federico Berto KAIST, Diff Eq ML fberto@kaist.ac.kr Stefano Massaroli The University of Tokyo, Diff Eq ML massaroli@robot.t.u-tokyo.ac.jp Michael Poli Stanford University, Diff Eq ML zymrael@cs.stanford.edu Jinkyoo Park KAIST jinkyoo.park@kaist.ac.kr |
| Pseudocode | No | The paper describes procedures and methods but does not include any formal pseudocode blocks or algorithms labeled as 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | We share the code used in this paper and make it publicly available on Github1. The following appendix also supplements the main text by providing additional clarifications. [...] 1Supporting reproducibility code is at https : //github.com/Diff Eq ML/diffeqml research/tree/master/hypersolvers control |
| Open Datasets | Yes | For Spring-Mass system: We select ξ(x, u) as a uniform distribution with support in X U where X = [ 20, 20] [ 20, 20] while U = [ 100, 100]. Nominal solutions are calculated on the accurate system using dopri5 with relative and absolute tolerances set to 10 7 and 10 9 respectively. (and similar descriptions for other systems with distributions ξ(x,u) and solvers). |
| Dataset Splits | No | The paper describes pre-training strategies for the hypersolvers and subsequent evaluation on optimal control tasks, but it does not specify explicit training/validation/test dataset splits or mention the use of a separate validation set for hyperparameter tuning or model selection in the typical machine learning sense. |
| Hardware Specification | Yes | Experiments were carried out on a machine equipped with an AMD RYZEN THREADRIPPER 3960X CPU with 48 threads and two NVIDIA RTX 3090 graphic cards. |
| Software Dependencies | No | The paper mentions key software libraries used (PyTorch, torchdyn, torchdiffeq) along with citations, but it does not provide explicit version numbers for these software components (e.g., 'PyTorch 1.9'). |
| Experiment Setup | Yes | We train the hypersolver on local residuals via stochastic exploration using the Adam optimizer with learning rate of 3 10 4 for 3 105 epochs.Table 1: Hyper parameters for the hypersolver networks in the experiments. |