# Sampling configs
samplers = [
pop.random_sampling,
pop.random_sampling,
pop.hv_improvement_sampling,
pop.max_mini_x_dist_sampling,
pop.max_mini_y_dist_sampling,
pop.hv_improvement_sampling,
]
sample_sizes = [1] * 6
sample_rates = [0.1, 0.8, 0.6, 0.6, 0.6, 1.]
sample_candidates = ['decision_space', 'pop'] + ['pm_hat'] * 3 + ['gap']
# ===============================Initialization================================
pop.config_surrogate(typ='rbf', params=params_surrogate, n_process=2)
pop.config_gap_opt(at='least_crowded',
radius=0.1,
size=POP_SIZE,
max_generation=MAX_GENERATION,
selection_fun=None,
mutation_fun=None,
mutation_rate=None,
crossover_fun=nsga_crossover,
trial_method='lhs',
trial_criterion='cm',
u=0.,
st=0.2)
#pop.config_sampling(methods=samplers, sizes=sample_sizes,
# sample_rates=sample_rates,
'batch_size': 4,
'warm_start': True,
'beta_1': 0.9,
'beta_2': 0.999,
'epsilon': 1e-12,
'alpha': 0.00001,
'learning_rate': 'adaptive',
'learning_rate_init': 0.002,
'max_iter': 500,
'verbose': False,
'no_improvement_tol': 500,
}
# ===============================Initialization================================
pop.config_surrogate(typ='ANN', params=params_surrogate)
pop.config_gap_opt(at='least_crowded',
radius=0.1,
size=2 * POP_SIZE,
max_generation=MAX_GENERATION,
selection_fun=None,
mutation_fun=None,
mutation_rate=None,
crossover_fun=nsga_crossover,
trial_method='lhs',
trial_criterion='cm',
u=0.,
st=0.02)
pop.config_sampling(methods='default',
sizes='default',
rate='default',
'trial_method': 'lhs',
'trial_criterion': 'cm'}
kernel = C(1.0, (1e-3, 1e3)) * RBF(1.0, (1e-3, 1e1)) + \
W(noise_level=1e-8, noise_level_bounds=(1e-10, 1e-5))
params_surrogate = \
{'kernel': kernel,
'alpha': 0.,
'optimizer': heuristic_optimizer,
'normalize_y': True,
}
# ===============================Initialization================================
pop.config_surrogate(typ='kriging', params=params_surrogate, n_process=2)
pop.config_gap_opt(at='least_crowded', radius=0.5, size=POP_SIZE,
max_generation=2*MAX_GENERATION, selection_fun=None,
mutation_fun=None, mutation_rate=None,
crossover_fun=None, trial_method='lhs',
trial_criterion='cm', u=0., st=0.12)
pop.config_sampling(methods='default', sizes='default', rate='default',
candidates='default')
pop.run(params_ea=params_ea, params_surrogate=params_surrogate, theo=theo)
# ================================Visualization================================
fig, (ax, ax_metric) = plt.subplots(1, 2, figsize=(10,4), dpi=100)
# Test surrogate results
final_arc = pop.true_front
'trial_method': 'lhs',
'trial_criterion': 'cm'
}
kernel = C(1.0, (1e-3, 1e3)) * RBF(1e-1, (1e-4, 1.)) + \
W(noise_level=1e-8, noise_level_bounds=(1e-10, 1e-5))
params_surrogate = \
{'kernel': kernel,
'alpha': 0.,
'normalize_y': True,
}
# ===============================Initialization================================
pop.config_surrogate(typ='kriging', params=params_surrogate)
pop.config_gap_opt(at='least_crowded',
radius=0.1,
size=2 * POP_SIZE,
max_generation=MAX_GENERATION,
selection_fun=None,
mutation_fun=None,
mutation_rate=None,
crossover_fun=random_crossover,
trial_method='lhs',
trial_criterion='cm',
u=0.,
st=0.02)
pop.config_sampling(methods='default',
sizes='default',
rate='default',
def run_sim(n, path, simtitle, REVOLUTION, PROBLEM, seed):
POP_SIZE = 100
MAX_GENERATION = 50
MAX_EPISODE = 100
MAX_EVAL = 1000
STOPPING_RULE = 'max_eval'
MUTATION_RATE = 0.1
MUTATION_U = 0.
MUTATION_ST = 0.2
REF = [-14.3, 0.05]
MINIMIZE = True
VERBOSE = True
X_SCALER = StandardScaler()
# Set global numpy random_state
np.random.seed(seed)
theo = np.genfromtxt('pareto_Kursawe.txt', dtype='f8')
# Instantiate a population
pop = GOMORS(size=POP_SIZE,
problem=PROBLEM,
max_generation=MAX_GENERATION,
max_episode=MAX_EPISODE,
reference=REF,
minimize=MINIMIZE,
stopping_rule=STOPPING_RULE,
max_eval=MAX_EVAL,
mutation_rate=MUTATION_RATE,
revolution=REVOLUTION,
embedded_ea=MOEAD,
verbose=VERBOSE,
no_improvement_step_tol=4)
pop.selection_fun = pop.compute_front
pop.mutation_fun = gaussian_mutator
pop.crossover_fun = random_crossover
# Parametrization
params_ea = {
'u': MUTATION_U,
'st': MUTATION_ST,
'trial_method': 'lhs',
'trial_criterion': 'cm'
}
kernel = CubicKernel
tail = LinearTail
params_surrogate = \
{'kernel': kernel,
'tail': tail,
'maxp': MAX_EVAL + POP_SIZE,
'eta': 1e-8,
}
# ===============================Initialization============================
pop.config_surrogate(typ='rbf',
params=params_surrogate,
n_process=1,
X_scaler=X_SCALER,
warm_start=True)
pop.config_gap_opt(at='least_crowded',
radius=0.1,
size=POP_SIZE,
max_generation=MAX_GENERATION,
selection_fun=None,
mutation_fun=None,
mutation_rate=None,
crossover_fun=random_crossover,
trial_method='lhs',
trial_criterion='cm',
u=0.,
st=0.2)
pop.config_sampling(methods='default',
sizes='default',
rate='default',
candidates='default')
pop.run(params_ea=params_ea, params_surrogate=params_surrogate, theo=theo)
# ============================= Save Results ================================ #
# path to save
directory = path + simtitle + '/' + str(n) + '/'
if not os.path.exists(directory):
os.makedirs(directory)
pop.render_features(pop.true_front).tofile(directory + 'xs.dat')
pop.render_targets(pop.true_front).tofile(directory + 'fs.dat')
np.array(pop.hypervol_diff).tofile(directory + 'hv_diff.dat')
np.array(pop.hypervol_cov).tofile(directory + 'hv_cov.dat')
np.array(pop.hypervol_index).tofile(directory + 'hv_ind.dat')
# ================================Visualization============================== #
# plot_res(pop=pop, ref=theo, directory=directory)
return pop
# Parametrization
params_ea = {'u': MUTATION_U,
'st': MUTATION_ST,
'trial_method': 'lhs',
'trial_criterion': 'cm'}
kernel = [CubicKernel, TPSKernel]
tail = LinearTail
params_surrogate = \
{'kernel': kernel,
'tail': tail,
'maxp': 2000,
'eta': 1e-8,
}
# ===============================Initialization================================
pop.config_surrogate(typ='rbf', params=params_surrogate)
pop.config_gap_opt(at='least_crowded', radius=0.5, size=POP_SIZE,
max_generation=2*MAX_GENERATION, selection_fun=None,
mutation_fun=None, mutation_rate=None,
crossover_fun=None, trial_method='lhs',
trial_criterion='cm', u=0., st=0.12)
pop.config_sampling(methods='default', sizes='default', rate='default',
candidates='default')
pop.surrogate.fit(b, y)
pop.surrogate.benchmark(bounds=np.repeat(pop.bounds, 3, axis=0), dim=3, obj_fun=kursawe)