[Dev] One source of truth per problem, and the equality of the modelings checked by computation

オープン
  • 何か問題がありますか？
#391 コメント 0 件 リアクション 0 件 担当者 1 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

@ocots がすでに取り組んでいます。

2026年11月6日 から。

評価

この issue はまだ評価されていません。

説明

Written by Claude (Claude Code) at the request of ocots, from a design discussion with ocots.

Goal

Each problem should have one source of truth for what is shared between modelings, and an automatic check for what is written more than once. Today a problem is written once per modeling (OptimalControl @def, its scalarized variant, JuMP), plus its metadata. Nothing checks that the dynamics, costs and constraints of two modelings are the same function. #371 compares NLPs, which happens after discretization and mixes the formulation with the transcription.

What we know

  • Declarative data (parameters, bounds, initial guess, reference values, names) can be shared directly: #377 (shared initial guess), #379 (declared fields), #376 (features computed from the model).
  • ProblemData (ext/JuMPModels/schemes/data.jl, #368, #369) already describes a problem by continuous functions (x, u, v, t), with no discretization in it. These functions are generic Julia code. They can be called with numbers, JuMP variables or symbolic variables, so the same data could feed a ModelingToolkit modeling (#308).
  • The @def models stay hand-written. They are the showcase of OptimalControl, and two independent writings make a better oracle than one generated from the other.
  • Exa is the only reason OptimalControl needs @def. The Exa model is generated from the syntax by CTParser (build_examodel). A model built from functions with the functional API of CTModels has no Exa builder. control-toolbox/CTDirect.jl#673 removes this limit: the components get a traceable form, and one transcription in CTDirect serves every modeler.

Proposal (to be discussed)

  1. Shared declarative data: one source per problem, read by every modeling (#377, #379).
  2. Equality of the functions, by computation: for every problem available in both OptimalControl and JuMP, evaluate the dynamics, the Lagrange and Mayer costs, and the path and boundary constraints of the @def model and of ProblemData at random points, and require equality. This checks the formulation before any discretization, without a hand-written list. Differences that are wanted (scaling with state_expressions, component names) must be recognised by computation and reported.
  3. ModelingToolkit (optional): build the MTK modeling from ProblemData instead of writing it by hand (#308).
  4. Target, after control-toolbox/CTDirect.jl#673: a ProblemData could become a CTModels Model with its traceable form, hence solvable with Exa without @def. Its transcription by CTDirect would then be compared with the hand-written JuMP transcription (schemes/build.jl), which stays an independent oracle.

Related

#257, #284, #308, #371, #376, #377, #379; control-toolbox/CTDirect.jl#673 and control-toolbox/CTDirect.jl#674.

How to work on this theme

This is a theme issue: it describes the work to do and what is known so far, it is not meant to be closed by a single PR.

  1. Whoever takes it first defines the work precisely and splits it into sub-issues (they can live in any repository of the organization).
  2. Work proceeds through the sub-issues; tick the checklist below as they are closed.
  3. Close the theme when everything is done or explicitly dropped.
Sub-issues
  • to be created when the theme is taken
主要言語
Julia
スター
13
フォーク
1
平均マージ
3時間 24分
マージ済み PR（30日）
34

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

control-toolbox/OptimalControlProblems.jl のほかの issue

control-toolbox/OptimalControlProblems.jl の issue をすべて見る

似ている issue

Julia の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。