StePPO: Adaptive Step-Size Control for Initial-Value Problems Using Proximal Policy Optimisation
Year: 2026
Type: software
Abstract: This capsule reproduces Figs. 2–5 of "StePPO: Adaptive Step-Size Control for IVPs Using PPO" from the released models, on CPU. Click Reproducible Run: it downloads the released StePPO models, the DeepONet baselines and the PID/Oracle datasets from Hugging Face, solves the test tasks with PID and StePPO, and writes each figure as a PDF, with the plotted numbers as CSVs, to /results. What this capsule does not do: it does not generate the datasets (it downloads the published ones), it does not train the models (StePPO training takes several GPU hours per system), and it runs without a GPU. Figs. 2–4 do not depend on the hardware; the Fig. 5 wall-clock times do, so they differ from the paper's GPU timings. To generate the datasets, train the models and time the solvers on a GPU yourself, see REPRODUCE.md in the GitHub repository (https://github.com/MKoumans/StePPO), section "Figures from scratch". It runs in the provided CUDA Docker container.
Source: Code Ocean
Institutions: Eindhoven University of Technology, ASML (Netherlands)
Language: English
Subfield: Statistical and Nonlinear Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214749294