Dataset and analysis scripts for: Towards Autonomous Bio-Inspired Optimization: Deep Reinforcement Learning for Adaptive Metaheuristic Orchestration

Dataset and analysis scripts Companion deposit for the article "Towards Autonomous Bio-Inspired Optimization: Deep Reinforcement Learning for Adaptive Metaheuristic Orchestration" (submitted to Biomimetics, MDPI; manuscript biomimetics-4483602). The article studies a deep reinforcement learning agent (PPO) that orchestrates a portfolio of seven bio-inspired, population-based metaheuristics over a single shared population, evaluated on the 0/1 multidimensional knapsack problem (Chu and Beasley cb9 set, n = 500, m = 30) under a budget of B = 25,000 evaluations with quantum Q = 25 and population N = 100. Contents cb9_instances.zip The thirty cb9 benchmark instances used in the study, transcribed from the official OR-Library file, one text file per instance (mknapcb9_instNN_ABIERTA.txt), together with: references_cb9.csv — best-known reference value per instance, taken from Lai, Hao, Glover and Lu (2018), "A two-phase tabu-evolutionary algorithm for the 0-1 multidimensional knapsack problem". orlib_official/mknapcb9.txt — the official OR-Library source file (J. E. Beasley, OR-Library), included verbatim so the transcription can be verified. orlib_official/mkcbres.txt — the official OR-Library results file. PROVENANCE.md — provenance notes for every file. results_evaluation.zip Raw evaluation results on the held-out test split (3 instances x 10 seeds = 30 runs per strategy, 12 strategies): one folder per seed with the per-run outputs of the three pipelines (base selection, joint selection-and-configuration, co-evolution), plus the consolidated files used in the article: comparison_todas_las_semillas.csv — every GAP behind the ranking and effect-size tables and the box-plot/heatmap figures. ranking_promedio.csv, rankings_por_semilla.xlsx, semillas_utilizadas.json, and the paper/ folder with the Wilcoxon statistics and exported figures. results_training.zip Training artifacts of the three PPO agents (selection, joint selection-and-configuration, co-evolution): per-episode logs, PPO update metrics, per-step decision logs, training summaries, validation and test evaluations of each run, and the final trained models (Stable-Baselines3 2.8). Intermediate checkpoints are omitted. The training-diagnostics and portfolio-usage figures of the article are generated from these logs. analysis_scripts.zip convert_orlib_mknapcb.py — converts the official OR-Library file into the per-instance format used here (reproduces the transcription). update_cb9_references.py — writes the best-known values of Lai et al. (2018) into the instance files (reproduces the reference table). eval_models_cb9.ipynb / eval_models_cb9_semillas_fijas.ipynb — the evaluation notebooks that produce the per-seed results and the consolidated ranking (the second one pins the ten test seeds used in the article). _make_training_figs.py, _make_results_figs.py — generate every results figure of the article from the CSV files in this deposit. Reproducibility notes All runs use a fixed instance split (seed 42, stratified by constraint tightness: 21 training / 6 validation / 3 test instances) and the seeds listed in semillas_utilizadas.json. Quality is reported as GAP = (f_ref - f) / f_ref x 100 against the best-known values of Lai et al. (2018). The evaluation always uses the complete repair operator. The scripts expect the folder layout of the research framework they belong to; they are included so that every number and figure in the article can be traced to raw data. The full framework is available from the corresponding author on request. Contact César Carrasco Carré — [e-mail] Escuela de Ingeniería Informática, Pontificia Universidad Católica de Valparaíso, Chile. How to cite Please cite the article once published; until then, cite this deposit via its DOI.