This dataset contains three summaries. These were generated through continuous optimization of the CEC2020 and SOCO2011 benchmarks, as well as for a real-world application involving portfolio optimization with the SP500 dataset.

The objective of these experiments is to analyze the performances and behaviors of a family of optimization algorithms, we named: 'fractal-based decomposition algorithms'. They hierarchically decompose a continuous search space using a self-similar and self-recurrent geometrical object. These algorithms can be described by five elementary building blocks: fractal, tree search, scoring, exploration and exploitation search components.

To obtain these experimental results, we used 11 different algorithms instantiated with our Python package named 'Zellij'. Each of these algorithms has a unique combination of building blocks. Thus, one can analyze the sensitivity to the five components, dimensionality, and problem definition.

The summaries are provided as CSV files. Each file contains 10 features (8 for summary_sp500.tab): the dimensionality of the benchmark function, the optimization algorithm, the benchmark function and basic statistics regarding the errors computed using the raw data (minimum, maximum, mean, standard error, median, first, and third quartile). Summaries are named according to the three experiments: summary_cec2020.tab, summary_socco2011.tab, summary_sp500.tab.

Raw data is provided as a compressed ZIP file, named raw_data.zip, it contains three folders for the three experiments. Each experiment folder contains subfolders for each algorithm, dimension, and benchmark function. These are organized as follows: /experiment_{name}/{algorithm}/D{dimension}_{function}_save/outputs/all_evaluations.csv.

A README.md file is provided for further technical details.

Python, 3.10.6

Pandas, 1.5.2

NumPy, 1.24.1

PyPortfolioOpt, 1.4.1

Zellij, 1.0.1