研究计划
组织平台证据层级的四个研究计划——范围、基准族、求解器阵容、方法学与论证门状态。
平台的证据被组织为四个研究计划,每个证据层级对应一个:核心、速度、质量与平台。 各层级层层递进:核心计划接纳核心与公开证据;速度计划再加入速度证据;质量计划再加入 质量证据;平台计划则呈现完整的公开领域。每个计划都精确界定其证据层级可揭示的内容, 因此较低层级的计划绝不会暴露较高层级的机制。
下方的展示区从已提交的计划数据包中读取每一份阵容、基准族、比较队列、方法学细节与 论证门状态,因此这些数据会随平台演进而保持最新,无需手工编辑文字。在某个研究计划的 论证门开启之前,不会发布任何比较性论断;当前发布的导出仅携带冒烟开发证据。
Four study programs organize the platform’s evidence tiers. Every roster, family list, methodology row, and claim-gate status below is build-inlined metadata from the committed catalog; no comparative claim is published until a program’s claim gate opens.
The foundational comparison tier over the baseline solver field.
Publication surface: Evidence-grade exports and the public portal's core catalog.
metadata-only
Each study program scopes what its evidence tier may reveal; no comparative claim is published until its claim gate opens.
Admitted evidence labels: core, public.
Claim gates
- gate closedfoundation track beats the comparison cohort with Holm-corrected significance before any superiority claim
No campaign at the statistical-power floor has produced disclosure-filtered comparative evidence on the public surface; published bundles carry smoke-development evidence only.
Benchmark families
Solver roster & comparison cohort
| Solver roster | Comparison cohort |
|---|---|
| ndso-core | ga-int(registry-internal) |
| — | brkga(registry-internal) |
| — | eda(registry-internal) |
| — | tabu-search(registry-internal) |
Methodology
- Objectives
- makespan
- Statistics
- wilcoxon-signed-rank, friedman-omnibus, holm-post-hoc, nemenyi-critical-difference, vargha-delaney-a12, cliffs-delta, bca-bootstrap-ci
- Run-count floor
- ≥30 independent runs per stochastic cell before comparative analysis.
- Campaign plan
- ablation → comparative → pilot → sensitivity → targeted
Planned figures & tables
Figures
- Convergence trajectory
- Solver rankings
- Critical-difference diagram
Tables
- Solver summary
- Pairwise comparisons
- Effect sizes
Evidence pointers
release-gated
A companion write-up is published only from an approved, disclosure-filtered evidence bundle after this program's claim gate and the release gates open.
The time-bounded comparison tier focused on runtime behavior.
Publication surface: Evidence-grade exports and the public portal's speed catalog.
metadata-only
Each study program scopes what its evidence tier may reveal; no comparative claim is published until its claim gate opens.
Admitted evidence labels: core, speed, public.
Claim gates
- gate closedspeed track reaches matched quality at a smaller budget than the comparison cohort under both stopping protocols
No campaign at the statistical-power floor has produced disclosure-filtered comparative evidence on the public surface; published bundles carry smoke-development evidence only.
Benchmark families
Solver roster & comparison cohort
| Solver roster | Comparison cohort |
|---|---|
| ndso-fast | iterated-greedy(registry-internal) |
| — | alns(registry-internal) |
| — | vns(registry-internal) |
Methodology
- Objectives
- makespan, lateness
- Statistics
- wilcoxon-signed-rank, friedman-omnibus, holm-post-hoc, nemenyi-critical-difference, vargha-delaney-a12, cliffs-delta, bca-bootstrap-ci
- Run-count floor
- ≥30 independent runs per stochastic cell before comparative analysis.
- Campaign plan
- ablation → comparative → pilot → sensitivity
Planned figures & tables
Figures
- Convergence trajectory
- Performance profile
- Solver rankings
Tables
- Solver summary
- Pairwise comparisons
- Budget speedup
Evidence pointers
release-gated
A companion write-up is published only from an approved, disclosure-filtered evidence bundle after this program's claim gate and the release gates open.
The solution-quality comparison tier focused on objective value.
Publication surface: Evidence-grade exports and the public portal's quality catalog.
metadata-only
Each study program scopes what its evidence tier may reveal; no comparative claim is published until its claim gate opens.
Admitted evidence labels: core, speed, quality, public.
Claim gates
- gate closedprecision track beats the comparison cohort on the primary quality indicator with Holm-corrected significance
No campaign at the statistical-power floor has produced disclosure-filtered comparative evidence on the public surface; published bundles carry smoke-development evidence only.
Benchmark families
Solver roster & comparison cohort
| Solver roster | Comparison cohort |
|---|---|
| ndso-summit | memetic(registry-internal) |
| — | nsga3 |
Methodology
- Objectives
- makespan, cost, composite
- Statistics
- wilcoxon-signed-rank, friedman-omnibus, holm-post-hoc, nemenyi-critical-difference, vargha-delaney-a12, cliffs-delta, bca-bootstrap-ci
- Run-count floor
- ≥30 independent runs per stochastic cell before comparative analysis.
- Campaign plan
- ablation → comparative → pilot → sensitivity → targeted
Planned figures & tables
Figures
- Convergence trajectory
- Pareto frontier
- Critical-difference diagram
Tables
- Solver summary
- Quality indicators
- Mechanism ablation
Evidence pointers
release-gated
A companion write-up is published only from an approved, disclosure-filtered evidence bundle after this program's claim gate and the release gates open.
The reproducibility and tooling tier over the full evidence set.
Publication surface: Evidence-grade exports, the public portal, and release notes.
metadata-only
Each study program scopes what its evidence tier may reveal; no comparative claim is published until its claim gate opens.
Admitted evidence labels: core, speed, quality, platform, public.
Claim gates
- gate closedplatform track routes any unsupported platform claim to limitations or future work rather than overclaiming
No campaign at the statistical-power floor has produced disclosure-filtered comparative evidence on the public surface; published bundles carry smoke-development evidence only.
- gate openCoverage and reproducibility statements are supported by artifact inspection.
platform-comparison.json feature rows are artifact-inspected.
Benchmark families
Solver roster & comparison cohort
| Solver roster | Comparison cohort |
|---|---|
| ndso-core | — |
| ndso-fast | — |
| ndso-summit | — |
Comparison basis: the rival-framework matrix published as platform-comparison.json.
Methodology
- Objectives
- makespan, cost
- Statistics
- feature-coverage, distribution-distance-bridge
- Run-count floor
- ≥30 independent runs per stochastic cell before comparative analysis.
- Campaign plan
- comparative → pilot
Planned figures & tables
Figures
- Feature-richness matrix
- Distribution-distance calibration
Tables
- Feature coverage
- Framework comparison
Evidence pointers
release-gated
A companion write-up is published only from an approved, disclosure-filtered evidence bundle after this program's claim gate and the release gates open.
Future program candidates
Candidate programs staged for a future evidence tier — each doubles as a worked example of the add-a-program recipe.
Constraint-and-robustness extension candidate
rcpspflow-shopDefer / retire rule: defer until a robustness-bearing campaign produces a labeled corpus; retire if no advantage clears the gate after one campaign
Top-level policy: Each study program scopes what its evidence tier may reveal; no comparative claim is published until its claim gate opens.
核心计划
核心计划研究基线求解器领域上的基础目标性能——这是支撑其他所有计划的比较。其基准族为 机器调度、作业车间与柔性作业车间,并以原生核心求解器组合对阵一支构造式与元启发式的 比较队列。其方法学结合 Wilcoxon 符号秩检验、带 Holm 事后校正的 Friedman 总体检验、 Nemenyi 临界差分分析,以及带 BCa 自助置信区间的 Vargha–Delaney A12 与 Cliff's delta 效应量,每个随机单元至少 30 次独立运行。在完整比较战役通过统计门之前,其比较论证门 保持关闭;当前发布的导出为冒烟开发证据。
速度计划
速度计划研究在匹配终止条件下的运行时与随时效率——随时曲线、性能剖面,以及在更小预算 下达到匹配质量的门。其基准族为流水车间与资源受限项目调度,其比较队列为迭代贪心、ALNS 与 VNS。它采用与核心计划相同的非参数方法学与运行下限。其比较论证门保持关闭,直到完整 比较战役通过统计门为止。
质量计划
质量计划研究解的质量与多目标指标——目标间隙、帕累托前沿与临界差分视图、质量指标表与 消融。其基准族为资源受限项目调度与柔性作业车间,其比较队列为模因搜索与 NSGA-III。它 沿用与上述计划相同的统计纪律与运行下限。其比较论证门保持关闭,直到完整比较战役通过 统计门为止。
平台计划
平台计划呈现平台所发布的最广领域——一份经过策展的跨族阵容,代表注册表中的每一个 求解器族,而非每一个已注册的求解器,包含特征覆盖、分布距离标签,以及可复现性与 工具证据。与平台检视视图一致,未获支持的论断会归入 局限而非过度宣称。其覆盖与可复现性态势如今由工件检视支撑,这是一种能力态势 而非性能论断;其比较门与其他计划一样保持关闭。
添加研究计划
新的研究计划遵循展示区为每个既有计划所呈现的同一条失败即关闭路径。展示区中的未来候选 行即为可参照的示例。
-
注册层级。 在层级注册表中添加一条层级条目及其证据层级取值;注册表契约测试会强制 完整性。
-
声明证据轨道。 向实验图集添加一条证据轨道——其基准族、求解器阵容、比较队列、 统计方法、论证门与运行下限——或先将其作为未来候选启动。
-
搭建工作区。 为该计划创建带配方的工作区目录,以及一个试点与一个比较战役配置:
New-Item -ItemType Directory -Path .\experiments\programs\<id> Copy-Item .\experiments\configs\core-pilot.json ` .\experiments\configs\<id>-pilot.json Copy-Item .\experiments\configs\core-comparative.json ` .\experiments\configs\<id>-comparative.json -
运行它。 为新层级执行试点战役,然后用 analytica 命令行工具整理分层数据包:
uv run python experiments\scripts\regenerate.py ` --config .\experiments\configs\<id>-pilot.json --tier <id> -
发布。 重新生成门户,该计划便会在此自动出现;若跳过重新生成,新鲜度门会使构建 失败:
uv run python tools\build_site_assets.py -
记录。 将该计划的章节添加到本页,并在各语言环境中同步。