ewarchul / cec2021

Results for the paper for Special Session and Competition on Real-Parameter Single Objective Optimization CEC 2021

CEC 2021

• About
• Content
  • Algorithms
  • Data
  • Benchmark configs
• Experiments reproduction
• Contact

About

This repository contains results of numerical experiments performed on CEC 2021 for the paper:

A new step-size adaptation rule for CMA-ES based on the population midpoint fitness

It also contains the source code of used optimization algorithms and scripts to reproduce conducted experiments.

Content

• R/ stores implementation of used algorithms in R
• data/ stores obtained data from numerical experiments:
  • cec13 has data from CEC 2013
  • cec17 has data from CEC 2017
  • cec21 has data from CEC 2021
• time-complexity contains R script to compute complexity of considered algorithm in the paper
• configs stores benchmark configuration files written in YAML.

Algorithms

Each algorithm is implemented in R programming language and based on the source code of the CRAN package {cmaesr}.

Data

Recorded data from experiments, i.e. error values (F(x) - F(x*)), is stored in m/ and M/ directories. The first directory contains the error values recorded with a budget step defined as in CEC 2021 specification:

d^(0:15/5 - 3) * maxFES

where d stands for the dimension and maxFES is budget for objective function evaluation.

The data from the directory M/ stores the error values recorded in the same manner as in CEC 2017 (or older):

c(0.01, 0.02, 0.03, 0.05, seq(0.1, 1, 0.1)) * maxFES

Benchmark configs

Experiments were performed using {cecb}, i.e. the R package written by me; mainly for the self-purposes. The package executes CECs experiments using the YAML data-serialization language. See the documentation for further details.

Also, we used {cecs} package which provides an interface from R to the benchmark functions implemented in C.

Experiments reproduction

If you want to reproduce our results you have to have installed docker in version 19.03.8 on your machine and run the commands written below in the root directory of the repository:

make build
make run cec=[CEC_VERSION | all]

These commands will build docker image and execute the bash script which takes options:

• CEC_VERSION which takes values from {2013, 2017, 2021} to reproduce experiments only for the one CEC version
• all which reproduces all experiments presented in the paper.

Contact

Feel free to contact me if you have any suggestions, questions, etc.: [e-mail]