!!! UNDER DEV WARNING !!!
This library is still under development.
All the examples are reproducible, but care should be taken when using this tool in production.
Before using this tool, one is ought to choose or define:
- The problem to be optimized
- Use a premade framework (recommended for the current stage)
- The MOEA
- The surrogate(s)
There are two recommended ways two construct a Problem:
-
Construct your problem with the
samolib._utils.construct_problemmethod.import numpy as np from samolib._utils import construct_problem # Write your objective functions, i.e. the ZDT-1 function def zdt1(x): f1 = x[0] g = np.add(1.0, np.multiply(9./29., np.sum(x[1:]))) f2 = np.multiply(g, (1. - np.sqrt(np.divide(f1, g)))) return np.array([f1, f2]) dim = 30 # Dimension of the problem n_objs = 2 # Number of objectives bounds = np.repeat([[0., 1.]], dim, axis=0) # Boundaries of the decision variables # Construct your problem my_problem = construct_problem(dim=dim, n_objs=n_objs, fun=zdt1, bounds=bounds)
-
Build your problem from scratch, using
samolib._utils.UnconstainedProblemas parent class:import numpy as np from samolib._utils import UnconstrainedProblem # Construct your problem class ZDT1(UnconstrainedProblem): def __init__(self): super(ZDT1, self).__init__(dim=30, n_objs=2, fun=zdt1) self.bounds = np.repeat([[0., 1.]], self.dim, axis=0) return None # The following function is optional, it computes the analytical optima of the ZDT-1 problem def solutions(self, resolution=1000): xs = np.zeros((resolution+1, self.dim)) xs[:, 0] = np.linspace(0., 1., resolution+1) return np.apply_along_axis(self.fun, 1, xs)
For the moment, we implemented two algorithm frameworks MOPRISM (Multi-objective Optimization by PRogressive Improvement Surrogate Model) and GOMORS (proposed in Akhtar and Shoemaker 2016 ). An example of how to instantiate a premade algorithm object is present in the next section.
This tool has implemented a naive MOEA (with random crossover operator, gaussian mutation operator for real-valued problems and random mutator for integer problems). For this option, you probably don't need to specify your option when declaring your algorithm.
The premade algorithm frameworks have implemented interfaces to communicate with MOEAs imported from the PYGMO library. PYGMO is implemented with C/C++, and runs relatively fast. A list of availble EAs in PYGMO can be found here. Since we are interested in MOPs, only NSGA-II and MOEAD are available in our toolbox to be chosen as embedded MOEA. Here is an example to declare MOPRISM with embedded MOEAD:
from samolib.ea import MOEAD
from samolib.premade import MOPRISM
from samolib.benchmarks import Kursawe
# Define problem
prob = Kursawe()
# Declare MOPRISM with MOEAD embedded
mop = MOPRISM(size=100, problem=prob, max_generation=50, reference=[-14, 0.],
minimize=True, stopping_rule='max_eval', max_eval=5000,
revolution=True, embedded_ea=MOEAD, local_search=True,
verbose=True, verbosity=5)For the moment, some ML models are available in this toolbox. The samolib.surrogates.NeuralNet, samolib.surrogates.Kriging, samolib.surrogates.SVM are implemented based on Scikit-Learn. The samolib.surrogates.RBFN is based on pySOT.
from sklearn.preprocessing import StandardScaler as X_SCALER
# Just apply sklearn.neural_network.MLPRegressor parameters here
params_surrogate = params_surrogate = \
{'hidden_layer_sizes': (6, 6),
'activation': 'logistic',
'solver': 'adam',
'early_stopping': False,
'batch_size': 32,
'warm_start': True,
'beta_1': 0.9,
'beta_2': 0.999,
'epsilon': 1e-12,
'alpha': 1e-3,
'learning_rate': 'constant',
'learning_rate_init': 0.0005,
'max_iter': 500,
'verbose': False,
'no_improvement_tol': 4,
}
# Configure the surrogate
mop.config_surrogate(typ='ANN', params=params_surrogate, n_process=2,
X_scaler=X_SCALER(), warm_start=True)import sys, os
import numpy as np
import itertools
sys.path.append('path_to_samolib/')
from premade import MOPRISM
from ea import MOEAD
from _utils import nsga_crossover, gaussian_mutator, random_crossover
from sklearn.preprocessing import StandardScaler
from benchmarks import ZDT1
# RBFN surrogate kernels
from pySOT.kernels import CubicKernel
from pySOT.tails import LinearTail
import matplotlib.pyplot as plt
from matplotlib import rc
# Matplotlib
rc('font', **{'family': 'sans-serif', 'sans-serif': ['Helvetica']})
# for Palatino and other serif fonts use:
# rc('font',**{'family':'serif','serif':['Palatino']})
rc('text', usetex=True)
PROBLEM = ZDT1()
POP_SIZE = 100
MAX_GENERATION = 50
MAX_EPISODE = 100
MAX_EVAL = 500
STOPPING_RULE = 'max_eval'
MUTATION_RATE = 0.1
MUTATION_U = 0.
MUTATION_ST = 0.2
REF = [1., 1.]
MINIMIZE = True
VERBOSE = True
LOCAL_SEARCH = True
X_SCALER = StandardScaler
# Set global numpy random_state
np.random.seed(0)
theo = PROBLEM.solutions()
# Instantiate a population
mop = MOPRISM(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=True,
embedded_ea=MOEAD, local_search=LOCAL_SEARCH, verbose=VERBOSE)
mop.selection_fun = mop.compute_front
mop.mutation_fun = gaussian_mutator
mop.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,
}
# Cofigure surrogate, local search and sampler
mop.config_surrogate(typ='rbf', params=params_surrogate, n_process=1,
X_scaler=X_SCALER(), warm_start=True)
mop.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)
mop.config_sampling(methods='default', sizes='default', rate='default',
candidates='default')
mop.run(params_ea=params_ea,
params_surrogate=params_surrogate,
theo=theo)# Do the same with GOMORS
from premade import GOMORS
# Set global numpy random_state
np.random.seed(0)
theo = PROBLEM.solutions()
# Instantiate a population
gom = GOMORS(size=POP_SIZE, problem=ZDT1(), max_generation=MAX_GENERATION,
max_episode=MAX_EPISODE, reference=REF, minimize=MINIMIZE,
stopping_rule=STOPPING_RULE, max_eval=MAX_EVAL,
mutation_rate=MUTATION_RATE, embedded_ea=MOEAD,
local_search=LOCAL_SEARCH, verbose=VERBOSE)
gom.selection_fun = gom.compute_front
gom.mutation_fun = gaussian_mutator
gom.crossover_fun = random_crossover
# Cofigure surrogate, local search and sampler
gom.config_surrogate(typ='rbf', params=params_surrogate, n_process=1,
X_scaler=X_SCALER(), warm_start=True)
gom.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)
gom.config_sampling(methods='default', sizes='default', rate='default',
candidates='default')
gom.run(params_ea=params_ea,
params_surrogate=params_surrogate,
theo=theo)ref = PROBLEM.solutions() # Compute analytical solutions
fs_mop = mop.render_targets('true_front') # Found objective values
hv_ind_mop = mop.hypervol_index # Number of evaluations where hypervolume was computed
hv_cov_effect_mop = mop.hypervol_cov_effect # Effective hypervolume (Only positive contributions counted)
fs_gom = gom.render_targets('true_front')
hv_ind_gom = gom.hypervol_index
hv_cov_effect_gom = gom.hypervol_cov_effect
fig, (ax, ax_metric) = plt.subplots(1, 2, figsize=(10, 4.5), dpi=100)
# Ranges
ax.set_xlim((0., 1.))
ax.set_ylim((0., 1.))
ax.set_xlabel(r'$f_1$')
ax.set_ylabel(r'$f_2$', labelpad=0)
ax.set(title="Non-dominated Solutions - %s" % PROBLEM.__class__.__name__)
# Plot
ax.plot(ref[:, 0], ref[:, 1], c='black', label="PS", zorder=100)
ax.scatter(fs_mop[:, 0], fs_mop[:, 1], c='orangered', s=16., label="MOPRISM", zorder=99)
ax.scatter(fs_gom[:, 0], fs_gom[:, 1], c='royalblue', s=20., label="GOMORS", marker='x')
# Plot legend
lgnd = ax.legend(loc='best', numpoints=1, fontsize=10)
# change the marker size manually for both lines
lgnd.legendHandles[0]._sizes = [10]
lgnd.legendHandles[1]._sizes = [16]
lgnd.legendHandles[2]._sizes = [16]
ax.grid(True, ls=':', which='both')
# ================================ Metrics ====================================
# Title
ax_metric.set_title('Hypervolume Coverage')
# Ranges
ax_metric.set_xlim((100, 500))
ax_metric.set_ylim((0., 1.))
# Lables
ax_metric.set_ylabel(r'Covergence \ Ratio', labelpad=None)
ax_metric.set_xlabel('Number of Evaluations (N)')
# Grid
ax_metric.grid(True, ls=':')
# Plot
hvplt_mop = ax_metric.plot(hv_ind_mop, hv_cov_effect_mop, label='MOPRISM', color='orangered',
linewidth=2.0)
hvplt_gom = ax_metric.plot(hv_ind_gom, hv_cov_effect_gom, label='GOMORS', color='royalblue',
linewidth=2.0)
# Ticks
ax_metric.set_yticklabels(ax_metric.get_yticks().round(2))
ax_metric.legend(loc='lower right', fontsize=10)
fig.tight_layout()
plt.show()