When Not to Reconfigure?

Language / 语言: English primary · 中文概览如下。

中文概览

研究非平稳需求和切换成本下的仓库重配置决策，回答何时应该延迟看似有利的布局调整。

Sequential warehouse reconfiguration under non-stationary demand and switching costs

Purpose: DECISION SCIENCE · Maturity: RESEARCH · Evidence: PUBLIC REAL-WORLD DATA

Part of TopPrism Decision Intelligence — Decision Science Research.

This repository is a research engine, not a production WMS or commercial slotting product.

Why this exists

Dynamic warehouse slotting systems are usually designed to answer how to improve a layout. This project asks a different question:

When should a warehouse deliberately defer a locally attractive physical reconfiguration?

A layout that looks better for the next demand period may still be a poor decision once relocation cost, future regime changes, forecast error and model error are considered.

The project studies how much of the full-information value of dynamic reconfiguration can actually be captured by deployable receding-horizon policies.

What this project studies

A stateful online reconfiguration problem — DWERP (Deferred Warehouse Reconfiguration Problem) — formalized inside the Metrical Task Systems / online optimization with switching costs family, and stress-tested against six public real-world demand cohorts.

World State S_t
     ↓
Candidate layout / routing experts
     ↓
Operating-cost estimate
+
Reconfiguration cost
     ↓
Sequential decision policy
     ↓
Layout L_t
     ↓
Execution / simulation
     ↓
State S_{t+1}

The repository also studies where real warehouse mechanics depart from a clean metric abstraction through capacity, exchange, batching and asymmetric operational effects.

Evidence

What is currently supported

• 6 public real-world demand datasets / 8 evaluation settings: WEPA, CrossStacks, Instacart (top-10% / mid-10%), Favorita, M5 (sparse / dense), SLAPRP.
• 29 reproducible experiments (R01–R29) under scripts/run_r*.py, with reports in outputs/experiments/.
• Three self-audit passes per major conclusion: fact check → inference check → method check. Withdrawn or downgraded claims are preserved in outputs/experiments/REVIEW_*.md.
• A 16-page, ~5,600-word, 25-reference manuscript (paper/main.pdf) with 5 figures, currently in academic-closure state (v1.5.5) ahead of submission.

What this evidence does not prove

• It does not establish that the approximate full-information planner is globally exact for all large instances. Small-instance exact certification is done on a limited grid (R18).
• It does not establish that all real warehouses exhibit zero deployable reconfiguration value. The empirical deployment-gap collapse (gap = 0.00% across the eight tested settings) is an observed result, not a universality proof.
• Public retail-demand datasets are not equivalent to full production WMS data with real geometry, equipment calibration and operational constraints.
• A seventh dataset (Footwear 2025) is referenced but could not be downloaded in the test environment; it is listed as future work (R30), not as completed evidence.
• The current project is not a production warehouse execution system.

Contribution map

The currently published contributions are three (see docs/PROGRESS_v1.5.5.md §2):

1. Structural opportunity exists. DWERP is formalized and stress-tested through seven T-gates (T0 diversity, T1a existence, T1b prevalence, T2 sensitivity, T1.5 move vs switch, T3 information boundary, T4 trap phase diagram). R18 shows beam-30 equals exact on 4/4 seeds; seed 17 reaches TrapScore 26.3.

2. The deployment paradox. Full-information reconfiguration opportunities exist, but their deployable value collapses under imperfect forecasts, imperfect internal cost models and realistic demand structure. Across the tested selector families, a fixed conservative policy (S1 FixedBest) remained a strong baseline against learned (S3 XGBoost, S4 MLP) and zero-shot LLM (S5) selectors.

3. Boundary and mechanism. Across six public datasets / eight evaluation settings, the empirical deployment-gap collapses to 0.00%. A clean MTS abstraction is systematically violated by real warehouse mechanics (asymmetric moves, capacity, batching, exchange chains); the gap between the theoretical metric d_m(L, L') and the implementation cost d_w(L, L') is itself a finding.

Where it fits at TopPrism

This project is part of TopPrism Decision Science Research. It explores reusable principles for stateful business decisions with transition costs.

Related TopPrism work:

• visit-scheduling-optimizer — periodic field-sales planning under recurring constraints.
• cultivating-ml-agent — project-driven capability accumulation for ML agents.

Status (v1.5.5)

Academic-closure state. The 29-run experiment set, three self-audit rounds, six public datasets and v4 manuscript are in place. Remaining items before first submission: R30 Footwear data acquisition attempt, a multi-seed CI on the main results, the cover letter and the supplementary paired per-instance comparison.

Productization would require additional validation with real warehouse geometry, operational cost calibration, production constraints and execution integration.

Repository structure

world_state/     state representation and public-data adapters
or_experts/      candidate optimization / routing experts
simulation/      execution and replay environment
features/        decision features
execution/       bounded execution / gateway logic
evaluation/      cost and policy evaluation
scripts/         reproducible experiment runners
outputs/         experiment reports and self-audit records
paper/           manuscript source and PDF
config/          experiment and cost configuration

Reproducibility

The repository keeps experiment outputs and self-audit records in version control so that positive findings, negative findings and withdrawn claims remain auditable.

Start from the latest progress document:

cat docs/PROGRESS_v1.5.5.md
cat docs/SPEC_UPDATE_v1.5.md

Then run the experiment script associated with the finding you want to reproduce under scripts/.

Research discipline

A distinctive project rule is that major conclusions are reviewed through three separate passes:

1. Fact check — verify results from code and artifacts rather than memory.
2. Inference check — verify that the stated conclusion is actually supported by the evidence.
3. Method check — record what the experiment taught us about the research process itself.

Withdrawn or downgraded conclusions are kept in the experiment history rather than silently rewritten.

Paper

See paper/main.pdf and paper/main.tex.

The current manuscript centers on the question:

When should a warehouse defer locally optimal physical reconfiguration under non-stationary demand, and how much of the full-information opportunity can deployable policies capture?

Boundaries & limitations

• empirical deployment-gap collapse is observed on the eight tested settings, not proved as a universal theorem.
• FI-Beam-30 remains an approximation; small-instance exact certification is reported only for the limited R18 grid.
• d_m ≠ d_w is acknowledged as a finding, not as a bug in the abstraction.
• The selector family comparison is reported on the tested datasets only; selector performance on a different deployment context has not been validated.
• The Footwear 2025 dataset is referenced but not included as completed evidence.
• Public retail-demand datasets are not production WMS data; conclusions do not transfer to live warehouse operations without additional validation.

License

MIT. See LICENSE.