def __init__(self, popsize=100, ksize=5):
# パラメータ
n_dim = 10
# popsize = 200
# epoch = 100
imax = 100 # 対象デブリ数
# ksize = 5
# 問題
problem = Problem(n_dim)
with Environment() as env:
# 個体クラス
indiv_type = Individual
# 初期個体生成クラス
indiv_pool = env.register(indiv_type)
# 遺伝子生成クラス
initializer = Initializer(n_dim, imax)
# GAオペレータ指定
# selection = TournamentSelection(ksize=2)
selection = TournamentSelectionStrict(ksize=2)
# selection = TournamentSelectionDCD()
# crossover = BlendCrossover(alpha=0.5)
crossover = Crossover(ix=OrderCrossover(rate=0.9),
dx=SimulatedBinaryCrossover(rate=0.9,
eta=20))
mutation = Mutation(im=SwapMutation(rate=0.1),
dm=PolynomialMutation(rate=0.1, eta=20))
ga_ops = {
'selection': selection,
'crossover': crossover,
'mutation': mutation
}
# optimizer = NSGA2(problem=problem, pool=indiv_pool, **ga_ops)
optimizer = MOEAD(problem=problem,
pool=indiv_pool,
ksize=5,
**ga_ops)
### Additional setting ###
Genome.set_weight([1, 3600] * n_dim)
Individual.set_weight([1, -1])
# optimizer.initializer = initializer
optimizer.n_cycle = popsize // 2
optimizer.alternation = 'all'
##########################
# 個体生成器
creator = Creator(initializer, indiv_pool)
# 登録
env.optimizer = optimizer
env.creator = creator
self.env = env
def main(model, out):
n_dim = 10
popsize = 30
epoch = 100 * 5
problem = Problem()
with Environment() as env:
indiv_pool = env.register(Individual)
indiv_pool.cls.set_weight([1, -1])
initializer = UniformInitializer(n_dim)
creator = Creator(initializer, indiv_pool)
crossover = SimulatedBinaryCrossover(rate=0.9, eta=100)
if model == 'moead':
ksize = 10
options = {
"ksize": ksize,
"normalization": True,
"crossover": crossover
}
optimizer = MOEAD(problem=problem, pool=indiv_pool, **options)
optimizer.weight_generator(nobj=4, divisions=50)
popsize = int(popsize)
epoch = epoch
elif model == 'nsga2':
optimizer = NSGA2(problem=problem,
pool=indiv_pool,
normalization=False)
elif model == 'para':
optimizer = NSGA2_para(problem=problem, pool=indiv_pool)
else:
raise Exception('Unexpected model name')
population = optimizer.init_population(creator, popsize=popsize)
history = [population]
print("obj weight:", population[0].data.weight)
for i in range(1, epoch + 1):
if i % 50 == 0:
print("epoch ", i)
population = optimizer(population)
history.append(population)
if i == epoch:
file = f'popsize{popsize}_epoch{epoch}_{ut.strnow("%Y%m%d_%H%M%S")}.pkl'
file = os.path.join(out, file)
if not os.path.exists(out):
os.makedirs(out)
print('save:', file)
ut.save(file, (env, optimizer, history))
return env, optimizer, history
def local_main():
n_dim = 30
popsize = 100
problem = zdt1
with Environment() as env:
# 個体クラス
indiv_type = Individual
# 初期個体生成クラス
indiv_pool = env.register(indiv_type)
# 遺伝子生成クラス
initializer = UniformInitializer(n_dim)
creator = Creator(initializer, indiv_pool)
# # 適応度クラス
# fit_type = Fitness
# # 初期個体生成クラス
# evaluator = env.register(fit_type)
###
# optimizer = NSGA2(problem=problem, pool=indiv_pool)
optimizer = MOEAD(problem=problem, pool=indiv_pool, ksize=5)
# optimizer.set_initializer(Initializer(3))
# optimizer.setup(problem)
### Additional setting ###
# optimizer.initializer = initializer
# optimizer.n_cycle = None
# optimizer.alternation = 'replace'
##########################
# indivs = [creator() for _ in range(popsize)]
# population = Population([ind.evaluate(problem) for ind in indivs])
# optimizer.calc_fitness(population)
population = optimizer.init_population(creator, popsize=popsize)
history = [population]
# print(population[0].data.origin)
# return
for i in range(100):
population = optimizer(population)
plot(population)
origin = population[0].data.origin.origin or []
# print([x.id for x in origin], '->', population[0].data.id)
history.append(population)
if i % 50 == 50 - 1:
ut.save(f'result/temp{i}.pkl', history)
plt.show()
return env, optimizer, history
def ga_result1(out='result'):
''' アニメーション表示
'''
file = ut.fsort(glob.glob(os.path.join(out, f'epoch*.pickle')))[-1]
optimizer = MOEAD.load(file=file)
print('epoch:', len(optimizer))
elite = optimizer.get_elite()
print('elite:', len(elite))
imax = len(optimizer) - 1
for epoch in range(0, imax, 20):
print(epoch, end='\n')
# plt.cla()
population = optimizer[epoch]
# population = optimizer.calc_rank(population)
front = [x for x in population]
x, y = np.array([x.data.value for x in front]).T
plt.scatter(x,
y,
color='blue',
alpha=epoch / imax,
label=f'epoch{epoch}')
# plt.legend()
plt.xlim((0, 50))
# plt.ylim((0, 50))
plt.pause(0.2)
plt.show()
def ga_main1(out='result'):
file = ut.fsort(glob.glob(os.path.join(out, f'epoch*.pickle')))[-1]
optimizer = MOEAD.load(file=file)
elite = optimizer.get_elite()
print('elite:', len(elite))
imax = len(optimizer)
for epoch, population in enumerate(optimizer):
print(epoch, end='\n')
plt.cla()
# print([fit.rank for fit in population])
# exit()
# for i in range(1):
# front = [x for x in population if x.rank == i]
# if not front:
# continue
# x, y = np.array([x.data.value for x in front]).T
# plt.scatter(x, y, label=f'{front[0][0]:.3f}')
front = population
x, y = np.array([x.data.value for x in front]).T
plt.scatter(x, y, label=f'{front[0][0]:.3f}')
# plt.legend()
# if epoch >= 10:
# plt.show()
# return
plt.pause(10 / imax)
plt.show()
def main1(out='result', force=False, clear_directory=False):
''' 最適化実行(ミニマル)
'''
''' ======== 最適化パラメータの定義 ======== '''
problem = zdt1
n_dim = 30
popsize = 100
epoch = 200
with Environment() as env:
''' ======== 必要なオブジェクトの準備 ======== '''
# 個体管理クラス
indiv_pool = env.register(Individual)
# 遺伝子生成クラス
initializer = UniformInitializer(n_dim)
# 初期個体生成クラス
creator = Creator(initializer, indiv_pool)
# 最適化オペレータ
op0 = NSGA2(problem=problem, pool=indiv_pool)
op1 = MOEAD(problem=problem, pool=indiv_pool, ksize=5)
optimizer = op0
''' ======== 初期集団の作成 ======== '''
population = optimizer.init_population(creator, popsize=popsize)
history = [population]
''' ======== 最適化 ======== '''
for i in range(epoch):
print('epoch', i + 1)
population = optimizer(population)
history.append(population)
if i % 50 == 50-1:
ut.save(f'result/temp{i}.pkl', history)
if i >= epoch // 2:
optimizer = op1
''' ======== 結果保存 ======== '''
data = env, optimizer, history
ut.save(f'result/optimize_#{ut.snow}.pkl', data)
return env, optimizer, history
def ga_main2(out='result'):
file = ut.fsort(glob.glob(os.path.join(out, f'epoch*.pickle')))[-1]
optimizer = MOEAD.load(file=file)
population = optimizer[-1]
front = [x for x in population if x.rank == 1]
front.sort(key=attrgetter('data.value'))
for ind in front:
print(ind.value, ind.data.value)
crowdings = [ind.value[1] for ind in front]
fig, axes = plt.subplots(2)
axes[0].plot(crowdings)
x, y = np.array([x.data.value for x in front]).T
im = axes[1].scatter(x, y, c=crowdings, cmap='jet')
# plt.xlim((0, 1))
# plt.ylim((0, 1))
fig.colorbar(im)
plt.show()
def __init__(self, problem):
# パラメータ
n_dim = 30
pop_size = 100
ksize = 5 # MOEA/D近傍サイズ
# epoch = 250
# save_trigger = lambda i: i == 1 or i % 10 == 0 # 10 epochごと
# save_trigger = lambda i: i == epoch # 最後だけ
if problem == zdt4 or problem == zdt6:
n_dim = 10
if problem == zdt4:
Individual.set_bounds([0.0] + [-5.0] * (n_dim - 1),
[1.0] + [5.0] * (n_dim - 1))
# Individual.set_bounds([0], [1])
# Individual.set_weight([1, 1])
# 問題
# problem = zdt4
initializer = UniformInitializer(n_dim)
selection = RandomSelection()
# selection = TournamentSelection(ksize=2)
# selection = TournamentSelectionStrict(ksize=2)
# selection = TournamentSelectionDCD()
# crossover = BlendCrossover(alpha=0.5)
crossover = SimulatedBinaryCrossover(rate=0.9, eta=20)
mutation = PolynomialMutation(rate=1/n_dim, eta=20)
optimizer = MOEAD(pop_size, selection, crossover, mutation, ksize)
# optimizer.set_initializer(Initializer(3))
optimizer.setup(problem)
### Additional setting ###
optimizer.initializer = initializer
optimizer.n_cycle = None
optimizer.alternation = 'all'
##########################
self.optimizer = optimizer
def __init__(self, pop_size=100, ksize=5):
# パラメータ
n_dim = 10
# pop_size = 200
# epoch = 100
imax = 100 # 対象デブリ数
# ksize = 5
# 問題
problem = Problem(n_dim)
initializer = Initializer(n_dim, imax)
# selection = TournamentSelection(ksize=2)
# selection = TournamentSelectionStrict(ksize=2)
selection = RandomSelection()
# selection = TournamentSelectionDCD()
# crossover = BlendCrossover(alpha=0.5)
crossover = Crossover(ix=OrderCrossover(rate=0.9),
dx=SimulatedBinaryCrossover(rate=0.9, eta=20))
mutation = Mutation(im=SwapMutation(rate=0.1),
dm=PolynomialMutation(rate=0.1, eta=20))
optimizer = MOEAD(pop_size, selection, crossover, mutation, ksize)
optimizer.setup(problem)
### Additional setting ###
Genome.set_weight([1, 7200] * n_dim)
Individual.set_weight([1, -1])
optimizer.initializer = initializer
optimizer.n_cycle = pop_size // 2
optimizer.alternation = 'new'
##########################
self.optimizer = optimizer