Running experiments

This is an example of an experimental study based on solving two problems of the ZDT family with two versions of the same algorithm (NSGA-II). The hypervolume indicator is used for performance assessment.

from jmetal.util.laboratory import Experiment, Job

def configure_experiment(problems: list, n_run: int):
 jobs = []

 for run in range(n_run):
     for problem in problems:
         jobs.append(
             Job(
                 algorithm=NSGAII(
                     problem=problem,
                     population_size=100,
                     mating_pool_size=100,
                     offspring_size=100,
                     max_evaluations=5000,
                     mutation=NullMutation(),
                     crossover=SBX(probability=1.0, distribution_index=20),
                     selection=BinaryTournamentSelection(comparator=RankingAndCrowdingDistanceComparator())
                 ),
                 label='NSGAII with Null Mutation',
                 run=run,
             )
         )
         jobs.append(
             Job(
                 algorithm=NSGAII(
                     problem=problem,
                     population_size=100,
                     mating_pool_size=100,
                     offspring_size=100,
                     max_evaluations=5000,
                     mutation=Polynomial(probability=1.0 / problem.number_of_variables, distribution_index=20),
                     crossover=SBX(probability=1.0, distribution_index=20),
                     selection=BinaryTournamentSelection(comparator=RankingAndCrowdingDistanceComparator())
                 ),
                 label='NSGAII with Polynomial Mutation',
                 run=run
             )
         )

 return jobs

if __name__ == '__main__':
 jobs = configure_experiment(problems=[ZDT1(), ZDT2()], n_run=3)

 test = Experiment(
     base_directory='./experiment',
     jobs=jobs
 )
 test.run()

 metrics = [HyperVolume(reference_point=[1, 1]), ComputingTime()]
 data = test.compute_metrics(metrics)
 print(data)

			NSGAII with Null Mutation	NSGAII with Polynomial Mutation
Problem	Metric	Run
ZDT1	Hypervolume	0	0.315708	0.516258
		1	0.323271	0.491973
		2	0.414953	0.507568
ZDT2	Hypervolume	0	0.293225	0.504027
		1	0.280499	0.417048
		2	0.357358	0.489576

The return value of compute_metrics() is a pandas DataFrame, so we can use all the methods that pandas makes available for us:

mean_results = data.groupby(['problem', 'metric']).mean()
median_results = data.groupby(['problem', 'metric']).median()
only_one_metric = data.xs('Hypervolume', level='metric')
min_values = data.groupby(['problem', 'metric']).min()
max_values = data.groupby(['problem', 'metric']).max()