Modular DE

This work-in-progress repository contains the code used to create a modular version of differential evolution.

Basic use-case: L-SHADE

To instantiate L-SHADE using modDE and optimize a function (using iohexperimenter), the following code can be used:

from modularde import ModularDE
import ioh
import numpy as np

f = ioh.get_problem(23, 1, 5)
lshade = ModularDE(f, base_sampler='uniform', mutation_base='target', mutation_reference='pbest', bound_correction='expc_center', crossover='bin', lpsr=True, lambda_ = 18*5, memory_size = 6, use_archive=True, init_stats=True, adaptation_method_F='shade', adaptation_method_CR='shade')
lshade.run()

To perform a larger benchmark experiment which includes tracking of internal parameters, the following can be used (note that running the full experiment with detailed tracking will use a significant amount of storage):

class LSHADE_interface():
    def __init__(self, bound_corr):
        self.bound_corr = bound_corr
        self.lshade = None
        
    def __call__(self, f):
        self.lshade = ModularDE(f, base_sampler='uniform', mutation_base='target', mutation_reference='pbest', bound_correction = self.bound_corr, crossover='bin', lpsr=True, lambda_ = 18*f.meta_data.n_variables, memory_size = 6, use_archive=True, init_stats = True, adaptation_method_F='shade', adaptation_method_CR='shade')
        self.lshade.run()
        
    @property
    def F(self):
        if self.lshade is None:
            return 0
        return self.lshade.parameters.stats.curr_F
    
    @property
    def CR(self):
        if self.lshade is None:
            return 0
        return self.lshade.parameters.stats.curr_CR

    @property
    def CS(self):
        if self.lshade is None:
            return 0
        return self.lshade.parameters.stats.CS
    
    @property
    def ED(self):
        if self.lshade is None:
            return 0
        return self.lshade.parameters.stats.ED
    
    @property
    def cumulative_corrected(self):
        if self.lshade is None:
            return 0
        return self.lshade.parameters.stats.corr_so_far
    
    @property
    def corrected(self):
        if self.lshade is None:
            return 0
        return self.lshade.parameters.stats.corrected
        
obj = LSHADE_interface('saturate')

exp = ioh.Experiment(algorithm = obj, #Set the optimization algorithm
  fids = range(1,25), iids = [1,2,3,4,5], dims = [5,30], reps = 5, problem_type = 'Real', #Problem definitions
  njobs = 12, logger_triggers = [ioh.logger.trigger.ALWAYS],#Enable paralellization
  logged = True, folder_name = f'L-SHADE_sat', algorithm_name = f'L-SHADE', store_positions = True, #Logging specifications
  experiment_attributes = {'SDIS' : 'Saturate'}, logged_attributes = ['corrected', 'cumulative_corrected', 'F', 'CR', 'CS', 'ED'], #Attribute tracking
  merge_output = True, zip_output = True, remove_data = True #Only keep data as a single zip-file
)

exp()

Note for version 0.0.4:

With the latest update, the internal seeding has been updated. However, the halton sampler's output depends on the used scipy version. Version 1.11 is highly recommended, but 1.10 also works but leads to slightly different behaviour for this sampler only.

Citation information

If you use this package, please consider to cite this paper:

@inproceedings{DBLP:conf/gecco/VermettenCKB23,
  author       = {Diederick Vermetten and
                  Fabio Caraffini and
                  Anna V. Kononova and
                  Thomas B{\"{a}}ck},
  editor       = {Sara Silva and
                  Lu{\'{\i}}s Paquete},
  title        = {Modular Differential Evolution},
  booktitle    = {Proceedings of the Genetic and Evolutionary Computation Conference,
                  {GECCO} 2023, Lisbon, Portugal, July 15-19, 2023},
  pages        = {864--872},
  publisher    = {{ACM}},
  year         = {2023},
  url          = {https://doi.org/10.1145/3583131.3590417},
  doi          = {10.1145/3583131.3590417},
  timestamp    = {Fri, 21 Jul 2023 22:25:47 +0200},
  biburl       = {https://dblp.org/rec/conf/gecco/VermettenCKB23.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

The design of this package is heavily based on the Modular CMA-ES package, created by Jacob de Nobel: https://github.com/IOHprofiler/ModularCMAES