Replication package for an integrated readiness-cost optimization framework

Description

This repository contains the replication package for a study of readiness–cost coordination in mission-critical supply chains. It examines how supplier selection, integer spare-parts inventories, maintenance scheduling, and crew capacity jointly affect system mission-capability probability and annual cost. The repository contains two ZIP archives. The dataset archive provides a curated public benchmark constructed from FAA Service Difficulty Reporting records for JASC Chapter 29 hydraulic systems, U.S. GAO readiness and sustainment reports, and USAspending PSC-1650 procurement records. FAA records inform failure-mode taxonomy, relative report composition, and occurrence-report diagnostics; they are not treated as exposure-normalized failure rates. USAspending data inform supplier-market archetypes and cost summaries. Supplier reliability, defect, lead-time, and geographic-risk attributes are transparent scenario parameters rather than observed causal estimates. The replication archive, mission_readiness_optimization.zip, contains the Python implementation, benchmark workbook, and final production-result snapshot. It includes the decision-dependent five-state continuous-time Markov chain, mixed-integer bi-objective optimization, pooled NSGA-II and MOEA/D reference-front construction, discrete-event simulation rank validation under exponential and Weibull failure times, corrected disruption re-optimization, response surfaces, and Sobol sensitivity analysis. Publication results are preserved in results/production/, while new runs are written to outputs/. Instructions are provided in the README. The workflow can be tested with python -u main.py --profile quick and reproduced with python -u main.py --profile production. The archived reference set pools 30 runs of each optimization algorithm and contains 587 non-dominated policies. Results show a substantial gap between the lowest-cost and highest-readiness observed policies, while knee and target policies recover much of the attainable mission-capability improvement at lower cost. Scheduling has the largest global total-order effect on readiness. Across fixed 80-90% mission-capability targets, disruptions primarily increase best-found cost. Maximum mission capability remains unchanged within a 0.1-percentage-point numerical tolerance under dominant-source outage, geopolitical shock, and broad supply degradation; doubled demand produces a 0.112-percentage-point reduction. All optimization results are best-found or observed solutions; global optimality is not claimed. The CTMC is a memoryless analytical surrogate, and DES evaluates policy-order preservation rather than absolute numerical agreement. Repository metadata and the dataset archive are released under CC BY 4.0. FAA, GAO, and USAspending records originate from public U.S. Government sources and remain subject to applicable source terms.

Steps to reproduce

Download both ZIP archives from this repository. Extract the dataset archive and the code archive. The code archive contains the Python implementation, while the dataset archive contains the public benchmark workbook and the associated documentation. Install Python 3.10 or later. From the root folder of the extracted code package, install the required dependencies using: pip install -r requirements.txt Run the complete reproducibility pipeline from the package root using: python main.py The script loads the benchmark workbook, validates the data contract, performs the empirical calibration, runs the decision-dependent Markov readiness model, approximates the readiness-cost Pareto frontier, executes the algorithm comparison, validates the analytical model through discrete-event simulation, runs the disruption scenarios, and computes the Sobol' sensitivity indices. Generated tables are written to outputs/tables/ and generated figures are written to outputs/figures/. For a faster smoke test, reduce the NSGA-II population size, number of generations, simulation replications, and Sobol' sample size in src/config.py. The full configuration used for the manuscript is documented in the README file.

Institutions

• University of Pisa

  Tuscany, Pisa
• University of Florence

  Tuscany, Florence
• Università degli Studi eCampus

  Lombardy, Novedrate
• University of Rome Tor Vergata

  Lazio, Rome

Categories

Licence