def __init__(self,
problem: Problem,
population_size: int,
cr: float,
f: float,
termination_criterion: TerminationCriterion,
k: float = 0.5,
population_generator: Generator = store.default_generator,
population_evaluator: Evaluator = store.default_evaluator,
dominance_comparator: Comparator = DominanceComparator()):
super(GDE3, self).__init__(problem=problem,
population_size=population_size,
offspring_population_size=population_size)
self.dominance_comparator = dominance_comparator
self.selection_operator = DifferentialEvolutionSelection()
self.crossover_operator = DifferentialEvolutionCrossover(cr, f, k)
self.population_generator = population_generator
self.population_evaluator = population_evaluator
self.termination_criterion = termination_criterion
self.observable.register(termination_criterion)
class GDE3(EvolutionaryAlgorithm[FloatSolution, FloatSolution]):
def __init__(self,
problem: Problem,
population_size: int,
cr: float,
f: float,
termination_criterion: TerminationCriterion = store.
default_termination_criteria,
k: float = 0.5,
population_generator: Generator = store.default_generator,
population_evaluator: Evaluator = store.default_evaluator,
dominance_comparator: Comparator = store.default_comparator):
super(GDE3, self).__init__(problem=problem,
population_size=population_size,
offspring_population_size=population_size)
self.dominance_comparator = dominance_comparator
self.selection_operator = DifferentialEvolutionSelection()
self.crossover_operator = DifferentialEvolutionCrossover(cr, f, k)
self.population_generator = population_generator
self.population_evaluator = population_evaluator
self.termination_criterion = termination_criterion
self.observable.register(termination_criterion)
def selection(self,
population: List[FloatSolution]) -> List[FloatSolution]:
mating_pool = []
for i in range(self.population_size):
self.selection_operator.set_index_to_exclude(i)
selected_solutions = self.selection_operator.execute(
self.solutions)
mating_pool = mating_pool + selected_solutions
return mating_pool
def reproduction(self, mating_pool: List[S]) -> List[S]:
offspring_population = []
first_parent_index = 0
for solution in self.solutions:
self.crossover_operator.current_individual = solution
parents = mating_pool[first_parent_index:first_parent_index + 3]
first_parent_index += 3
offspring_population.append(
self.crossover_operator.execute(parents)[0])
return offspring_population
def replacement(
self, population: List[S], offspring_population: List[FloatSolution]
) -> List[List[FloatSolution]]:
tmp_list = []
for solution1, solution2 in zip(self.solutions, offspring_population):
result = self.dominance_comparator.compare(solution1, solution2)
if result == -1:
tmp_list.append(solution1)
elif result == 1:
tmp_list.append(solution2)
else:
tmp_list.append(solution1)
tmp_list.append(solution2)
join_population = population + offspring_population
return RankingAndCrowdingDistanceSelection(
self.population_size,
dominance_comparator=self.dominance_comparator).execute(
join_population)
def create_initial_solutions(self) -> List[FloatSolution]:
return [
self.population_generator.new(self.problem)
for _ in range(self.population_size)
]
def evaluate(self,
solution_list: List[FloatSolution]) -> List[FloatSolution]:
return self.population_evaluator.evaluate(solution_list, self.problem)
def stopping_condition_is_met(self) -> bool:
return self.termination_criterion.is_met
def get_result(self) -> List[FloatSolution]:
return self.solutions
def get_name(self) -> str:
return 'GDE3'
from jmetal.util.solution_list import read_solutions, print_function_values_to_file, print_variables_to_file
from jmetal.util.termination_criterion import StoppingByEvaluations
from jmetal.util.visualization import Plot, InteractivePlot
if __name__ == '__main__':
problem = DTLZ2()
problem.reference_front = read_solutions(
filename='../../resources/reference_front/DTLZ2.3D.pf'.format(
problem.get_name()))
max_evaluations = 150000
algorithm = MOEAD(
problem=problem,
population_size=300,
crossover=DifferentialEvolutionCrossover(CR=1.0, F=0.5, K=0.5),
mutation=PolynomialMutation(probability=1.0 /
problem.number_of_variables,
distribution_index=20),
aggregative_function=Tschebycheff(
dimension=problem.number_of_objectives),
neighbor_size=20,
neighbourhood_selection_probability=0.9,
max_number_of_replaced_solutions=2,
weight_files_path='../../resources/MOEAD_weights',
termination_criterion=StoppingByEvaluations(max=max_evaluations))
algorithm.observable.register(observer=ProgressBarObserver(
max=max_evaluations))
algorithm.observable.register(observer=VisualizerObserver(
reference_front=problem.reference_front, display_frequency=1000))
def run(self) -> List[S]:
pool_1_size = self.population_size
pool_2_size = self.population_size
selection_operator_1 = BinaryTournamentSelection()
crossover_operator_1 = IntegerSBXCrossover(1.0, 20.0)
mutation_operator_1 = IntegerPolynomialMutation(
1.0 / self.problem.number_of_variables, 20.0)
selection_operator_2 = DifferentialEvolutionSelection()
crossover_operator_2 = DifferentialEvolutionCrossover(0.2, 0.5, 0.5)
dominance = DominanceComparator()
max_iterations = self.max_iterations
iterations = 0
parent_1: List[IntegerSolution] = [None, None]
generational_hv: List[float] = []
current_gen = 0
"""Create the initial subpopulation pools and evaluate them"""
pool_1: List[IntegerSolution] = []
for i in range(pool_1_size):
pool_1.append(self.problem.create_solution())
pool_1[i] = self.problem.evaluate(pool_1[i])
pool_2: List[IntegerSolution] = []
for i in range(pool_2_size):
pool_2.append(self.problem.create_solution())
pool_2[i] = self.problem.evaluate(pool_2[i])
evaluations = pool_1_size + pool_2_size
mix = self.mix_interval
problem = self.problem
h = HyperVolume(reference_point=[1] *
self.problem.number_of_objectives)
initial_population = True
"""The main evolutionary cycle"""
while iterations < max_iterations:
combi: List[IntegerSolution] = []
if not initial_population:
offspring_pop_1: List[IntegerSolution] = []
offspring_pop_2: List[IntegerSolution] = []
"""Evolve pool 1"""
for i in range(pool_1_size):
parent_1[0] = selection_operator_1.execute(pool_1)
parent_1[1] = selection_operator_1.execute(pool_1)
child_1: IntegerSolution = crossover_operator_1.execute(
parent_1)[0]
child_1 = mutation_operator_1.execute(child_1)
child_1 = problem.evaluate(child_1)
evaluations += 1
offspring_pop_1.append(child_1)
"""Evolve pool 2"""
for i in range(pool_2_size):
parent_2: List[
IntegerSolution] = selection_operator_2.execute(pool_2)
crossover_operator_2.current_individual = pool_2[i]
child_2 = crossover_operator_2.execute(parent_2)
child_2 = problem.evaluate(child_2[0])
evaluations += 1
result = dominance.compare(pool_2[i], child_2)
if result == -1:
offspring_pop_2.append(pool_2[i])
elif result == 1:
offspring_pop_2.append(child_2)
else:
offspring_pop_2.append(child_2)
offspring_pop_2.append(pool_2[i])
ind_1 = pool_1[random.randint(0, pool_1_size - 1)]
ind_2 = pool_2[random.randint(0, pool_2_size - 1)]
offspring_pop_1.append(ind_1)
offspring_pop_2.append(ind_2)
offspring_pop_1.extend(pool_1)
pool_1 = self.r.replace(offspring_pop_1[:pool_1_size],
offspring_pop_1[pool_1_size:])
pool_2 = self.r.replace(offspring_pop_2[:pool_2_size],
offspring_pop_2[pool_2_size:])
mix -= 1
if mix == 0:
"""Time to perform fitness sharing"""
mix = self.mix_interval
combi = combi + pool_1 + pool_2
# print("Combi size: ", len(combi))
"""pool1size/10"""
combi = self.r.replace(
combi[:int(pool_1_size / 10)],
combi[int(pool_1_size / 10):len(combi)],
)
"""
print(
"Sizes: ",
len(pool_1) + len(combi),
len(pool_2) + len(combi),
"\n",
)
"""
pool_1 = self.r.replace(pool_1, combi)
pool_2 = self.r.replace(pool_2, combi)
if initial_population:
initial_population = False
iterations += 1
print("Iterations: ", str(iterations))
"""
hval_1 = h.compute([s.objectives for s in pool_1])
hval_2 = h.compute([s.objectives for s in pool_2])
print("hval_1: ", str(hval_1))
print("hval_2: ", str(hval_2), "\n")
"""
new_gen = int(evaluations / self.report_interval)
if new_gen > current_gen:
combi = combi + pool_1 + pool_2
combi = self.r.replace(combi[:(2 * pool_1_size)],
combi[(2 * pool_1_size):])
hval = h.compute([s.objectives for s in combi])
for i in range(current_gen, new_gen, 1):
generational_hv.append(hval)
current_gen = new_gen
"""#Write runtime generational HV to file"""
"""Return the first non dominated front"""
combi_ini: List[IntegerSolution] = []
combi_ini.extend(pool_1)
combi_ini.extend(pool_2)
combi_ini = self.r.replace(
combi_ini[:pool_1_size + pool_2_size],
combi_ini[pool_1_size + pool_2_size:],
)
return combi_ini
def run(self) -> List[S]:
# selection operator 1
selection_operator_1 = BinaryTournamentSelection()
# selection operator 2
selection_operator_2 = DifferentialEvolutionSelection()
# crossover operator 1
crossover_operator_1 = SBXCrossover(1.0, 20.0)
# crossover operator 2
crossover_operator_2 = DifferentialEvolutionCrossover(0.2, 0.5, 0.5)
# crossover operator 3
crossover_operator_3 = DifferentialEvolutionCrossover(1.0, 0.5, 0.5)
# mutation operator 1
mutation_operator_1 = PolynomialMutation(
1.0 / self.problem.number_of_variables, 20.0)
# dominance comparator
dominance = DominanceComparator()
# array that stores the "generational" HV quality
generational_hv: List[float] = []
parent_1: List[FloatSolution] = [None, None]
parent_2: List[FloatSolution] = []
parent_3: List[FloatSolution] = []
# initialize some local and global variables
pool_1: List[FloatSolution] = []
pool_2: List[FloatSolution] = []
# size of elite subset used for fitness sharing between subpopulations
nrOfDirectionalSolutionsToEvolve = int(self.population_size / 5)
# subpopulation 1
pool_1_size = int(self.population_size -
(nrOfDirectionalSolutionsToEvolve / 2))
# subpopulation 2
pool_2_size = int(self.population_size -
(nrOfDirectionalSolutionsToEvolve / 2))
print(
str(pool_1_size) + " - " + str(nrOfDirectionalSolutionsToEvolve) +
" - " + str(self.mix_interval))
evaluations = 0
current_gen = 0
directionalArchiveSize = 2 * self.population_size
weights = self.__create_uniform_weights(
directionalArchiveSize, self.problem.number_of_objectives)
directionalArchive = self.__create_directional_archive(weights)
neighbourhoods = self.__create_neighbourhoods(directionalArchive,
self.population_size)
nrOfReplacements = 1
iniID = 0
# Create the initial pools
# pool1
pool_1: List[FloatSolution] = []
for _ in range(pool_1_size):
new_solution = self.problem.create_solution()
new_solution = self.problem.evaluate(new_solution)
evaluations += 1
pool_1.append(new_solution)
self.__update_extreme_values(new_solution)
dr = directionalArchive[iniID]
dr.curr_sol = new_solution
iniID += 1
# pool2
pool_2: List[FloatSolution] = []
for _ in range(pool_2_size):
new_solution = self.problem.create_solution()
new_solution = self.problem.evaluate(new_solution)
evaluations += 1
pool_2.append(new_solution)
self.__update_extreme_values(new_solution)
dr = directionalArchive[iniID]
dr.curr_sol = new_solution
iniID += 1
# directional archive initialization
pool_A: List[FloatSolution] = []
while iniID < directionalArchiveSize:
new_solution = self.problem.create_solution()
new_solution = self.problem.evaluate(new_solution)
evaluations += 1
pool_A.append(new_solution)
self.__update_extreme_values(new_solution)
dr = directionalArchive[iniID]
dr.curr_sol = new_solution
iniID += 1
mix = self.mix_interval
h = HyperVolume(reference_point=[1] *
self.problem.number_of_objectives)
insertionRate: List[float] = [0, 0, 0]
bonusEvals: List[int] = [0, 0, nrOfDirectionalSolutionsToEvolve]
testRun = True
# record the generational HV of the initial population
combiAll: List[FloatSolution] = []
cGen = int(evaluations / self.report_interval)
if cGen > 0:
combiAll = pool_1 + pool_2 + pool_A
combiAll = self.r.replace(
combiAll[:pool_1_size + pool_2_size],
combiAll[pool_1_size + pool_2_size:],
)
hval = h.compute([s.objectives for s in combiAll])
for _ in range(cGen):
generational_hv.append(hval)
current_gen = cGen
# the main loop of the algorithm
while evaluations < self.max_evaluations:
offspringPop1: List[FloatSolution] = []
offspringPop2: List[FloatSolution] = []
offspringPop3: List[FloatSolution] = []
dirInsertPool1: List[FloatSolution] = []
dirInsertPool2: List[FloatSolution] = []
dirInsertPool3: List[FloatSolution] = []
# evolve pool1 - using SPEA2 evolutionary model
nfe: int = 0
while nfe < (pool_1_size + bonusEvals[0]):
parent_1[0] = selection_operator_1.execute(pool_1)
parent_1[1] = selection_operator_1.execute(pool_1)
child1a: FloatSolution = crossover_operator_1.execute(
parent_1)[0]
child1a = mutation_operator_1.execute(child1a)
child1a = self.problem.evaluate(child1a)
evaluations += 1
nfe += 1
offspringPop1.append(child1a)
dirInsertPool1.append(child1a)
# evolve pool2 - using DEMO SP evolutionary model
i: int = 0
unselectedIDs: List[int] = []
for ID in range(len(pool_2)):
unselectedIDs.append(ID)
nfe = 0
while nfe < (pool_2_size + bonusEvals[1]):
index = random.randint(0, len(unselectedIDs) - 1)
i = unselectedIDs[index]
unselectedIDs.pop(index)
parent_2 = selection_operator_2.execute(pool_2)
crossover_operator_2.current_individual = pool_2[i]
child2 = crossover_operator_2.execute(parent_2)
child2 = self.problem.evaluate(child2[0])
evaluations += 1
nfe += 1
result = dominance.compare(pool_2[i], child2)
if result == -1: # solution i dominates child
offspringPop2.append(pool_2[i])
elif result == 1: # child dominates
offspringPop2.append(child2)
else: # the two solutions are non-dominated
offspringPop2.append(child2)
offspringPop2.append(pool_2[i])
dirInsertPool2.append(child2)
if len(unselectedIDs) == 0:
for ID in range(len(pool_2)):
unselectedIDs.append(random.randint(
0,
len(pool_2) - 1))
# evolve pool3 - Directional Decomposition DE/rand/1/bin
IDs = self.__compute_neighbourhood_Nfe_since_last_update(
neighbourhoods, directionalArchive,
nrOfDirectionalSolutionsToEvolve)
nfe = 0
for j in range(len(IDs)):
if nfe < bonusEvals[2]:
nfe += 1
else:
break
cID = IDs[j]
chosenSol: FloatSolution = None
if directionalArchive[cID].curr_sol != None:
chosenSol = directionalArchive[cID].curr_sol
else:
chosenSol = pool_1[0]
print("error!")
parent_3: List[FloatSolution] = [None, None, None]
r1 = random.randint(0, len(neighbourhoods[cID]) - 1)
r2 = random.randint(0, len(neighbourhoods[cID]) - 1)
r3 = random.randint(0, len(neighbourhoods[cID]) - 1)
while r2 == r1:
r2 = random.randint(0, len(neighbourhoods[cID]) - 1)
while r3 == r1 or r3 == r2:
r3 = random.randint(0, len(neighbourhoods[cID]) - 1)
parent_3[0] = directionalArchive[r1].curr_sol
parent_3[1] = directionalArchive[r2].curr_sol
parent_3[2] = directionalArchive[r3].curr_sol
crossover_operator_3.current_individual = chosenSol
child3 = crossover_operator_3.execute(parent_3)[0]
child3 = mutation_operator_1.execute(child3)
child3 = self.problem.evaluate(child3)
evaluations += 1
dirInsertPool3.append(child3)
# compute directional improvements
# pool1
improvements = 0
for j in range(len(dirInsertPool1)):
testSol = dirInsertPool1[j]
self.__update_extreme_values(testSol)
improvements += self.__update_neighbourhoods(
directionalArchive, testSol, nrOfReplacements)
insertionRate[0] += (1.0 * improvements) / len(dirInsertPool1)
# pool2
improvements = 0
for j in range(len(dirInsertPool2)):
testSol = dirInsertPool2[j]
self.__update_extreme_values(testSol)
improvements += self.__update_neighbourhoods(
directionalArchive, testSol, nrOfReplacements)
insertionRate[1] += (1.0 * improvements) / len(dirInsertPool2)
# pool3
improvements = 0
for j in range(len(dirInsertPool3)):
testSol = dirInsertPool3[j]
self.__update_extreme_values(testSol)
improvements += self.__update_neighbourhoods(
directionalArchive, testSol, nrOfReplacements)
# on java, dividing a floating number by 0, returns NaN
# on python, dividing a floating number by 0, returns an exception
if len(dirInsertPool3) == 0:
insertionRate[2] = None
else:
insertionRate[2] += (1.0 * improvements) / len(dirInsertPool3)
for dr in directionalArchive:
offspringPop3.append(dr.curr_sol)
offspringPop1 = offspringPop1 + pool_1
pool_1 = self.r.replace(offspringPop1[:pool_1_size],
offspringPop1[pool_1_size:])
pool_2 = self.r.replace(offspringPop2[:pool_2_size],
offspringPop2[pool_2_size:])
combi: List[FloatSolution] = []
mix -= 1
if mix == 0:
mix = self.mix_interval
combi = combi + pool_1 + pool_2 + offspringPop3
print("Combi size: " + str(len(combi)))
combi = self.r.replace(
combi[:nrOfDirectionalSolutionsToEvolve],
combi[nrOfDirectionalSolutionsToEvolve:],
)
insertionRate[0] /= self.mix_interval
insertionRate[1] /= self.mix_interval
if insertionRate[2] != None:
insertionRate[2] /= self.mix_interval
"""
print(
"Insertion rates: "
+ str(insertionRate[0])
+ " - "
+ str(insertionRate[1])
+ " - "
+ str(insertionRate[2])
+ " - Test run:"
+ str(testRun)
)
"""
if testRun:
if (insertionRate[0] > insertionRate[1]) and (
insertionRate[0] > insertionRate[2]):
print("SPEA2 win - bonus run!")
bonusEvals[0] = nrOfDirectionalSolutionsToEvolve
bonusEvals[1] = 0
bonusEvals[2] = 0
if (insertionRate[1] > insertionRate[0]) and (
insertionRate[1] > insertionRate[2]):
print("DE win - bonus run!")
bonusEvals[0] = 0
bonusEvals[1] = nrOfDirectionalSolutionsToEvolve
bonusEvals[2] = 0
if (insertionRate[2] > insertionRate[0]) and (
insertionRate[2] > insertionRate[1]):
print("Directional win - no bonus!")
bonusEvals[0] = 0
bonusEvals[1] = 0
bonusEvals[2] = nrOfDirectionalSolutionsToEvolve
else:
print("Test run - no bonus!")
bonusEvals[0] = 0
bonusEvals[1] = 0
bonusEvals[2] = nrOfDirectionalSolutionsToEvolve
testRun = not testRun
insertionRate[0] = 0.0
insertionRate[1] = 0.0
insertionRate[2] = 0.0
pool_1 = pool_1 + combi
pool_2 = pool_2 + combi
print("Sizes: " + str(len(pool_1)) + " " + str(len(pool_2)))
pool_1 = self.r.replace(pool_1[:pool_1_size],
pool_1[pool_1_size:])
pool_2 = self.r.replace(pool_2[:pool_2_size],
pool_2[pool_2_size:])
self.__clear_Nfe_history(directionalArchive)
hVal1 = h.compute([s.objectives for s in pool_1])
hVal2 = h.compute([s.objectives for s in pool_2])
hVal3 = h.compute([s.objectives for s in offspringPop3])
newGen = int(evaluations / self.report_interval)
if newGen > current_gen:
print("Hypervolume: " + str(newGen) + " - " + str(hVal1) +
" - " + str(hVal2) + " - " + str(hVal3))
combi = combi + pool_1 + pool_2 + offspringPop3
combi = self.r.replace(combi[:self.population_size * 2],
combi[self.population_size * 2:])
hval = h.compute([s.objectives for s in combi])
for j in range(current_gen, newGen):
generational_hv.append(hval)
current_gen = newGen
# return the final combined non-dominated set of maximum size = (populationSize * 2)
combiAll: List[FloatSolution] = []
combiAll = combiAll + pool_1 + pool_2 + pool_A
combiAll = self.r.replace(combiAll[:self.population_size * 2],
combiAll[self.population_size * 2:])
return combiAll
def get_algorithm_instance(algo_name):
algos = {
'smpso':
SMPSO(problem=objective_function,
swarm_size=swarm_size,
mutation=PolynomialMutation(probability=mutation_probability,
distribution_index=20),
leaders=CrowdingDistanceArchive(100),
termination_criterion=StoppingByEvaluations(
max_evaluations=max_evaluations)),
'omopso':
OMOPSO(problem=objective_function,
swarm_size=swarm_size,
epsilon=0.0075,
uniform_mutation=UniformMutation(
probability=mutation_probability, perturbation=0.5),
non_uniform_mutation=NonUniformMutation(
mutation_probability,
perturbation=0.5,
max_iterations=int(max_evaluations / swarm_size)),
leaders=CrowdingDistanceArchive(100),
termination_criterion=StoppingByEvaluations(
max_evaluations=max_evaluations)),
'nsgaii':
NSGAII(problem=objective_function,
population_size=30,
offspring_population_size=30,
mutation=PolynomialMutation(probability=mutation_probability,
distribution_index=20),
crossover=SBXCrossover(probability=1.0, distribution_index=20),
termination_criterion=StoppingByEvaluations(
max_evaluations=max_evaluations)),
'spea2':
SPEA2(problem=objective_function,
population_size=30,
offspring_population_size=30,
mutation=PolynomialMutation(probability=mutation_probability,
distribution_index=20),
crossover=SBXCrossover(probability=1.0, distribution_index=20),
termination_criterion=StoppingByEvaluations(
max_evaluations=max_evaluations)),
'moead':
MOEAD(
problem=objective_function,
population_size=30,
crossover=DifferentialEvolutionCrossover(CR=1.0, F=0.5, K=0.5),
mutation=PolynomialMutation(probability=mutation_probability,
distribution_index=20),
aggregative_function=Tschebycheff(
dimension=objective_function.number_of_objectives),
neighbor_size=5,
neighbourhood_selection_probability=0.9,
max_number_of_replaced_solutions=2,
weight_files_path='resources/MOEAD_weights',
termination_criterion=StoppingByEvaluations(max_evaluations=700)),
'ibea':
IBEA(problem=objective_function,
kappa=1.0,
population_size=30,
offspring_population_size=30,
mutation=PolynomialMutation(probability=mutation_probability,
distribution_index=20),
crossover=SBXCrossover(probability=1.0, distribution_index=20),
termination_criterion=StoppingByEvaluations(max_evaluations))
}
return algos[algo_name]