Experimental Evaluation of the NSGA-III on the 3-dimensional OneMinMax benchmark

This repository contains the source code and data used in the experimental evaluation of the NSGA-III on the 3-dimensional OneMinMax benchmark. The results are used in our paper "A Mathematical Runtime Analysis of the Non-dominated Sorting Genetic Algorithm III (NSGA-III)" accpeted for IJCAI 2023.

Files

• code_experiments.ipynb: A jupyter notebook containing the code for the benchmarking, writing the resulting data into .csv files and visualizing data from these .csv files

• exp_nsga3.csv: A file containing the data generated by benchmarking the NSGA-III on 3-OneMinMax tracking the number of iterations until the complete Pareto front is sampled. Each data point consists of the used algorithm (algo, always NSGA-3), the block size (blockSize) which is half the length of the bit strings in the population, the number of employed divisions along each objective for the reference points (refPoints), the population size (popSize), the number of iterations until the complete Pareto front was sampled (iterations) and the chance with which crossover was applied in the reproduction step (ratioCO). The column ratioMutate (corresponding to 1 - ratioCO is deprecated and can be ignored).

• exp_coverage.csv: A file containing the data generated by benchmarking the NSGA-III and NSGA-II on 3-OneMinMax tracking the ratio of the covered Pareto front for each generation of the optimization process. Each data point consists of the benchmark problem, always 3OMM = 3-OneMinMax), the used algorithm (algo, either nsga2 or nsga3), the block size (blockSize) which is half the length of the bit strings in the population, the number of employed divisions along each objective for the reference points (refPoints) for the NSGA-3 (0 for NSGA-2), the population size (popSize), the chance with which crossover was applied in the reproduction step (ratioCO), the number of tracked iterations (until either the maximum iteration, here 500, was hit or the complete Pareto front was found) and coverages, a list stating for each generation the number of elements in the Pareto front contained in the population of that generation.