RefineGreed — Reproducibility Package

This repository is the reproducibility package for the paper "Scalable RefineGreed: A Targeted Greedy-Refinement Heuristic for Large-Scale Assignment Problems" (submitted to the European Journal of Operational Research, May 2026). Version 2 extends the original dataset with revised algorithm implementations, an operationally-motivated case study dataset, and full experimental results. DATASETS Version 1 — Synthetic Benchmark (original): Benchmark instances for the Linear Assignment Problem (LAP) covering three cost matrix types — Uniform, Structured, and Skewed — with sizes from 10×10 up to 9000×9000 and 30 independently generated instances per size. Files are named type_size_instance.csv (e.g., uniform_100_1.csv). Each file is a plain n×n integer CSV without headers. Generated using LAP_Generator.py for full reproducibility. Version 2 — Operationally-Motivated Case Study (new): 40 cost matrices with log-normal cost structure (μ=2.7, σ=0.6 in log-space) calibrated to urban delivery travel times, spanning n = 100 to 800 with 5 instances per size. Files are named LAP_n_instance.csv and provided in CaseStudy_40_RawDatasets.zip. All instances are fully reproducible using fixed seeds (seed = n × 100 + instance). Distribution calibrated following Simonetto et al. (2019), DOI: 10.1016/j.trc.2019.01.019. ALGORITHMS Seven Python algorithm files are included: Hungarian Algorithm (HA_Auto.py), Auction Algorithm (AA_Auto.py), Standard Greedy Algorithm (SGA_Auto.py), and the proposed RefineGreed (RG_Auto.py). Three baseline algorithms have been revised from Version 1: Iterated Greedy (IG_Auto_Revised.py) now uses cost-guided greedy reconstruction; Simulated Annealing (SA_Auto_Revised.py) uses a correctly computed cooling schedule spanning the full run; Genetic Algorithm (GA_Auto_Revised.py) uses a population and generation budget scaled with problem size n. Full parameter details are documented in Table 4b of the paper. RESULTS Table7_YourAlgorithms_FullResults.csv contains full instance-level costs, execution times, and RPD values for all 7 algorithms on all 40 case study matrices. RPD = (Algorithm Cost − HA Optimal Cost) / HA Optimal Cost × 100. REQUIREMENTS: Python 3.8+, numpy, pandas, scipy AUTHORS: Yash Kumar ([e-mail]), Ramesh Chandra Sahoo, Prashant Dixit