Fitness-based Conditional RV-GOMEA

Code associated with the GECCO 2024 research paper titled: "Fitness-based Linkage Learning and Maximum-Clique Conditional Linkage Modelling for Gray-box Optimization with RV-GOMEA", by Georgios Andreadis, Tanja Alderliesten, and Peter A.N. Bosman. This codebase includes contributions by Anton Bouter and Chantal Olieman.

Building

Run make in this directory to build the project. The Armadillo library needs to be installed on the system. Depending on your installation, please modify the Makefile to link to it, correctly.

Running

To start a single run of the algorithm, run the command below. This will give you an overview of options with which you can control the algorithm.

python rvgomea/cmd/run.py --help

To reproduce the experiments in the pre-print, you can use the scripts in the run_scripts folder. SLURM configurations for each are available in the deploy and deploy_shark folders. Due to the volume of runs needed for the bisection scalability experiments, there are also job array scripts available. These divide the bulk of runs needed into one job per run.

An example run, launching the first 500 runs with at most 40 in parallel, would look as follows:

sbatch --array=1-500%40 deploy/parallel_scalability_thin.slurm

At times, when runs are stopped pre-maturely, it can help to clean-up left-over convergence files:

find data/scalability-all-results/ -name 'statistics.dat' -delete

To send only aggregate results (excluding individual results) to a different location for the extrapolation runs, use the following command:

cd gecco-data && find . -type f -name 'aggregated_results.csv' | tar -cf - -T - | tar -xf - -C scalability-aggregated

Testing

To run a quick test of different linkage models, try the following command:

make && run_scripts/test_all_models

To test the VkD-CMA implementation on the different problems, run:

run_scripts/test_all_models_with_vkd_cma

License

Fitness-based Conditional Real-Valued Gene-pool Optimal Mixing Evolutionary Algorithm © 2024 by Georgios Andreadis, Tanja Alderliesten, Peter A.N. Bosman, Anton Bouter, and Chantal Olieman is licensed under CC BY-NC-ND 4.0. A copy of the license is included in the LICENSE file.