Overview
The collection holds 22 problems. Each one has a page, generated from its files:
Problem: its description and mathematical formulation, its parameters, and its reference objective;
one tab per modeling where it exists: its status, how to get and solve it, the model as written, its results, and its solutions drawn;
Compare, when it exists in two modelings or more: which modelings are checked to give the same NLP, their results side by side, which solution each one reached, and the solutions of the modelings drawn together.
The tab bar stays at the top of the page, and the tab you choose is kept from one problem to the next. A link can open a page on a tab: add ?view=JuMP to its address.
All the problems shows them as cards, with the shape of their solution and their status in each modeling, or as a table to sort and filter by their features; the Status page shows every problem in every modeling, with the latest convergence reports.
What each problem comes with
| Where it comes from | |
|---|---|
| Modelings | the files of the problem: a problem exists in a modeling when its file does |
| Status | the metadata, valid or draft per modeling |
| Description | ext/MetaData/<problem>.md, next to the metadata |
| Parameters, default grid, initial guess | the metadata, ext/MetaData/<problem>.jl |
| Features | computed from the OptimalControl model |
| Reference objective | computed from the convergence and robustness crons, per scheme, on the default grid |
| Analytic solution | the metadata, when the continuous problem has a known solution |
The solutions drawn
The solutions on the pages are those of the weekly convergence crons: each solve records its trajectory (time grid, state, control, costate and variable, on at most 501 nodes), and your browser draws it. Choose the solves to overlay, the components, and the view: against time, in a phase plane, or in 3D (drag to rotate); move the time cursor, or play it, and every view follows. Draw up to t draws the trajectories up to the cursor only.
Each modeling says whether its results are up to date: solved for the current version of the problem, the one of the page. A problem is identified by a fingerprint of its files (its metadata, description and models): when one of them changes, the results are marked stale until the next run of the cron. The Reproducibility block gives the run, the commit, and the environment of each platform: Julia, system, processor, the versions of the packages, and the Project.toml and Manifest.toml to download.
The JuMP costate is the multiplier of the dynamics constraints, with the sign convention of OptimalControl (-JuMP.dual, the MOI ConstraintDual negated: #326).
The reference objective
For each scheme, the reference objective is the best objective among the solves that converged, for every modeling and configuration, on the default grid of the problem. The crons solve every problem every week in every configuration, and a script gathers the results into ext/MetaData/references/<problem>.toml; see how the references are computed.
The same file keeps every result and the distinct solutions the solves reached: two objectives closer than a relative OptimalControlProblems.SAME_SOLUTION are the same solution. A problem with several solutions has local solutions, and its Compare tab says which configuration reached which.
The reference is the objective of the discretized problem. When the continuous problem has a known solution, its objective is on the page too, and the gap between the two is the discretization error.
Results are reported, not tested
The tests of the package fail when a problem does not build or does not run, or when two modelings do not give the same NLP. Convergence is not tested: every solve is shown with its gap to the reference, as it is, without tolerance.
The same configuration does not reach exactly the same objective on every platform: Linux, Windows, macOS, Intel and ARM processors, and GPUs round floating-point operations differently, and a solver stops at a point within its tolerance. The results of each problem page give the platforms of each configuration and, when they differ, the range of their objectives; a solve on your machine lands in that range, or close to it.