Benchmark data for MSC-CMA-ES on CEC2014, CEC2017, CEC2020 and CEC2022

Raw and aggregated benchmark data supporting the article "MSC-CMA-ES: Structure-Aware Restarts for CMA-ES via Cyclic Nearest-Better Basin Discovery" (Nedanovski, Nenov, Pilev). Four suites, one zstd-compressed tar archive each: CEC2014 (D = 10, 30), CEC2017 (D = 10, 30), CEC2020 (D = 5, 10, 15, 20), CEC2022 (D = 10, 20). Seven algorithms — MSC-CMA-ES together with BIPOP-CMA-ES, ARRDE, j2020, jSO, L-SRTDE and NL-SHADE-RSP — at 51 runs per function. Each cell is run at the suite’s official evaluation budget and at several extended budgets; across the collection these range from 5×104 to 6×107 function evaluations. Layout inside every archive: <suite>/d<D>/<algorithm>/maxevals_<N>/, holding one .pkl per function with the per-run records and a summary.csv with the per-function aggregates. Per-cell and per-suite READMEs and the generated figures are included. Algorithm directory names use the internal short forms MSC-CMA, BIPOP-CMA, LSRTDE and NLSHADE-RSP. The .pkl files are the primary data: every quantity reported in the article — mean, median and best final error, and fixed-budget target coverage over 51 log-uniform targets in [1e+2, 1e-8] — is recomputable from them. The summary.csv files are those aggregates as used in the article’s tables and figures. Two-dimensional cells and exploratory CEC2019 runs are excluded; neither enters the article. Code: https://github.com/snenovgmailcom/cma_es_project, commit a888416; a snapshot of that commit is included here as source.tar.gz. SHA256SUMS covers the four data archives. Preprint: arXiv:2606.15830.