def apply(population_size: int, individual_size: int, bounds: np.ndarray,
func: Callable[[np.ndarray], float], opts: Any,
callback: Callable[[Dict], Any], max_evals: int, seed: Union[int,
None],
population: Union[np.array, None], answer: Union[None, float, int]):
"""
Applies the Self-adaptive differential evolution algorithm (SaDE).
:param population_size: Size of the population.
:type population_size: int
:param individual_size: Number of gens/features of an individual.
:type individual_size: int
:param bounds: Numpy ndarray with individual_size rows and 2 columns.
First column represents the minimum value for the row feature.
Second column represent the maximum value for the row feature.
:type bounds: np.ndarray
:param func: Evaluation function. The function used must receive one
parameter.This parameter will be a numpy array representing an individual.
:type func: Callable[[np.ndarray], float]
:param opts: Optional parameters for the fitness function.
:type opts: Any type.
:param callback: Optional function that allows read access to the state of all variables once each generation.
:type callback: Callable[[Dict], Any]
:param max_evals: Number of evaluatios after the algorithm is stopped.
:type max_evals: int
:param seed: Random number generation seed. Fix a number to reproduce the
same results in later experiments.
:type seed: Union[int, None]
:return: A pair with the best solution found and its fitness.
:rtype [np.ndarray, int]
"""
np.random.seed(seed)
if population is None:
population = commons.init_population(population_size, individual_size,
bounds)
fitness = commons.apply_fitness(population, func, opts)
num_evals = 0
pf1, pf2, pf3 = 1 / 3, 1 / 3, 1 / 3
cr1, cr2, cr3 = 1 / 3, 1 / 3, 1 / 3
bin_cross = 0.5
best_mutation = 0.5
current_generation = 1
learning_period = 50
update = True
s = collections.defaultdict(lambda: np.zeros(learning_period))
learning_k = 0
run_mmts = True
mmts_desired_evals = 60
num_no_mmts = 0
while num_evals < max_evals:
# 1. Generate ensemble parameters
if current_generation == 1 or current_generation > learning_period:
f = np.random.choice([0.3, 0.5, 0.7],
p=[pf1, pf2, pf3],
size=population_size)
cr = np.random.choice([0.1, 0.5, 0.9],
p=[cr1, cr2, cr3],
size=population_size)
mutation_strat = np.random.choice(
['rand', 'best'],
p=[1 - best_mutation, best_mutation],
size=population_size)
cross_method = np.random.choice(['bin', 'exp'],
p=[bin_cross, 1 - bin_cross],
size=population_size)
# 2.0 Niching
m = 5
pop_b = population.copy().reshape(population.shape[0], 1,
population.shape[1])
distances = np.sqrt(
np.einsum('ijk, ijk->ij', population - pop_b, population - pop_b))
neighbors = np.argsort(distances, axis=1)
l_best_indexes = neighbors[:, 1]
neighbors = neighbors[l_best_indexes, 1:m]
# 2.1 Mutation
rand_mut_idx = np.where(mutation_strat == 'rand')[0]
best_mut_idx = np.where(mutation_strat == 'best')[0]
mutated = np.empty(population.shape)
# 2.1.a Rand mutation
# Generate parents indexes
choices = np.arange(0, neighbors.shape[1])
parents1 = np.empty(population.shape)
parents2 = np.empty(population.shape)
parents3 = np.empty(population.shape)
for i in range(population_size):
choice = np.random.choice(choices, 3, replace=False)
parents1[i] = population[neighbors[i, choice[0]]]
parents2[i] = population[neighbors[i, choice[1]]]
parents3[i] = population[neighbors[i, choice[2]]]
mutated[rand_mut_idx] = parents1[
rand_mut_idx] + f[rand_mut_idx].reshape(len(rand_mut_idx), 1) * (
parents2[rand_mut_idx] - parents3[rand_mut_idx])
mutated[best_mut_idx] = population[
l_best_indexes[best_mut_idx]] + f[best_mut_idx].reshape(
len(best_mut_idx),
1) * (parents1[best_mut_idx] - parents2[best_mut_idx])
mutated = commons.keep_bounds(mutated, bounds)
# 2.2 Crossover
bin_cross_idx = np.where(cross_method == 'bin')[0]
exp_cross_idx = np.where(cross_method == 'exp')[0]
crossed = np.empty(population.shape)
crossed[bin_cross_idx] = commons.crossover(
population[bin_cross_idx], mutated[bin_cross_idx],
cr[bin_cross_idx].reshape(len(cr[bin_cross_idx]), 1))
crossed[exp_cross_idx] = commons.exponential_crossover(
population[exp_cross_idx], mutated[exp_cross_idx],
cr[exp_cross_idx].reshape(len(cr[exp_cross_idx]), 1))
# 2.3 Recalculate fitness
c_fitness = commons.apply_fitness(crossed, func, opts)
num_evals += population_size
# 2.4 Distance between new population and original population
distances = np.sqrt(
np.einsum('ijk, ijk->ij', crossed - pop_b, crossed - pop_b))
neighbors = np.argsort(distances, axis=1)
l_best_indexes = neighbors[:, 1]
selection = [
c_fitness[i] < fitness[l_best_indexes[i]]
for i in range(len(population))
]
population[l_best_indexes[selection]] = crossed[selection]
fitness[l_best_indexes[selection]] = c_fitness[selection]
if 1 < current_generation < learning_period:
num_no_mmts += 60
if num_no_mmts >= num_evals_mmts:
run_mmts = True
num_no_mmts = 0
if current_generation > learning_period:
if np.random.randn() > p_mmts:
run_mmts = True
if run_mmts:
selected = clearing(0.2, 5, population, fitness.copy())
a = mmts.mmts(population[selected], bounds, fitness[selected],
mmts_desired_evals, func, opts)
population[selected] = a[0]
fitness_test = fitness[selected] < a[1]
s['mmts_ok'] = len(fitness_test)
s['mmts_fail'] = 5 - len(fitness_test)
fitness[selected] = a[1]
num_evals += a[2]
if current_generation < learning_period:
num_evals_mmts = a[2]
run_mmts = False
# 3. Update control parameters
s['pf1_ok'][learning_k] = sum(np.logical_and(f == 0.3, selection))
s['pf1_fail'][learning_k] = sum(f == 0.3) - s['pf1_ok'][learning_k]
s['pf2_ok'][learning_k] = sum(np.logical_and(f == 0.5, selection))
s['pf2_fail'][learning_k] = sum(f == 0.5) - s['pf2_ok'][learning_k]
s['pf3_ok'][learning_k] = sum(np.logical_and(f == 0.7, selection))
s['pf3_fail'][learning_k] = sum(f == 0.7) - s['pf3_ok'][learning_k]
s['cr1_ok'][learning_k] = sum(np.logical_and(cr == 0.1, selection))
s['cr1_fail'][learning_k] = sum(cr == 0.1) - s['cr1_ok'][learning_k]
s['cr2_ok'][learning_k] = sum(np.logical_and(cr == 0.5, selection))
s['cr2_fail'][learning_k] = sum(cr == 0.5) - s['cr2_ok'][learning_k]
s['cr3_ok'][learning_k] = sum(np.logical_and(cr == 0.9, selection))
s['cr3_fail'][learning_k] = sum(cr == 0.9) - s['cr3_ok'][learning_k]
selection = np.array(selection)
s['rand_ok'][learning_k] = sum(selection[rand_mut_idx])
s['rand_fail'][learning_k] = len(
rand_mut_idx) - s['rand_ok'][learning_k]
s['best_ok'][learning_k] = sum(selection[best_mut_idx])
s['best_fail'][learning_k] = len(
best_mut_idx) - s['best_ok'][learning_k]
s['bin_ok'][learning_k] = sum(selection[bin_cross_idx])
s['bin_fail'][learning_k] = len(
bin_cross_idx) - s['bin_ok'][learning_k]
s['exp_ok'][learning_k] = sum(selection[exp_cross_idx])
s['exp_fail'][learning_k] = len(
exp_cross_idx) - s['exp_ok'][learning_k]
s['saepsde_ok'][learning_k] = len(selection)
s['saepsde_fail'][learning_k] = population_size - len(selection)
learning_k = (learning_k + 1) % learning_period
current_generation += 1
if current_generation > learning_period:
sf1 = np.sum(s['pf1_ok']) / (np.sum(s['pf1_ok']) +
np.sum(s['pf1_fail'])) + 0.02
sf2 = np.sum(s['pf2_ok']) / (np.sum(s['pf2_ok']) +
np.sum(s['pf2_fail'])) + 0.02
sf3 = np.sum(s['pf3_ok']) / (np.sum(s['pf3_ok']) +
np.sum(s['pf3_fail'])) + 0.02
pf1 = sf1 / (sf1 + sf2 + sf3)
pf2 = sf2 / (sf1 + sf2 + sf3)
pf3 = sf3 / (sf1 + sf2 + sf3)
pcr1 = np.sum(s['cr1_ok']) / (np.sum(s['cr1_ok']) +
np.sum(s['cr1_fail'])) + 0.02
pcr2 = np.sum(s['cr2_ok']) / (np.sum(s['cr2_ok']) +
np.sum(s['cr2_fail'])) + 0.02
pcr3 = np.sum(s['cr3_ok']) / (np.sum(s['cr3_ok']) +
np.sum(s['cr3_fail'])) + 0.02
cr1 = pcr1 / (pcr1 + pcr2 + pcr3)
cr2 = pcr2 / (pcr1 + pcr2 + pcr3)
cr3 = pcr3 / (pcr1 + pcr2 + pcr3)
pmut_best = np.sum(s['best_ok']) / (np.sum(s['best_ok']) +
np.sum(s['best_fail'])) + 0.02
pmut_rand = np.sum(s['rand_ok']) / (np.sum(s['rand_ok']) +
np.sum(s['rand_fail'])) + 0.02
best_mutation = pmut_best / (pmut_best + pmut_rand)
pcross_bin = np.sum(s['bin_ok']) / (np.sum(s['bin_ok']) +
np.sum(s['bin_fail'])) + 0.02
pcross_exp = np.sum(s['exp_ok']) / (np.sum(s['exp_ok']) +
np.sum(s['exp_fail'])) + 0.02
bin_cross = pcross_bin / (pcross_bin + pcross_exp)
p_saepsde = np.sum(s['saepsde_ok']) / (
np.sum(s['saepsde_ok']) + np.sum(s['saepsde_fail'])) + 0.02
p_mmts = np.sum(s['mmts_ok']) / (np.sum(s['mmts_ok']) +
np.sum(s['mmts_fail'])) + 0.02
p_mmts = p_mmts / (p_saepsde + p_mmts)
if callback is not None:
callback(**(locals()))
best = np.argmin(fitness)
if fitness[best] == answer:
yield population[best], fitness[best], population
break
else:
yield population[best], fitness[best], population
def apply(population_size: int, individual_size: int, bounds: np.ndarray,
func: Callable[[np.ndarray],
float], opts: Any, callback: Callable[[Dict], Any],
lambdas: Union[list, np.array], ng: int, c: Union[int, float],
p: Union[int, float], max_evals: int, seed: Union[int, None],
population: Union[np.array,
None], answer: Union[None, float,
int]) -> [np.ndarray, int]:
"""
Applies the MPEDE differential evolution algorithm.
:param population_size: Size of the population (NP-max)
:type population_size: int
:param ng: Number of generations after the best strategy is updated.
:type ng: int
:param lambdas: Percentages of each of the 4 subpopulations.
:type lambdas: Union[list, np.array]
:param individual_size: Number of gens/features of an individual.
:type individual_size: int
:param bounds: Numpy ndarray with individual_size rows and 2 columns.
First column represents the minimum value for the row feature.
Second column represent the maximum value for the row feature.
:type bounds: np.ndarray
:param func: Evaluation function. The function used must receive one
parameter.This parameter will be a numpy array representing an individual.
:type func: Callable[[np.ndarray], float]
:param opts: Optional parameters for the fitness function.
:type opts: Any type.
:param callback: Optional function that allows read access to the state of all variables once each generation.
:type callback: Callable[[Dict], Any]
:param max_evals: Number of evaluations after the algorithm is stopped.
:type max_evals: int
:param seed: Random number generation seed. Fix a number to reproduce the
same results in later experiments.
:param p: Parameter to choose the best vectors. Must be in (0, 1].
:type p: Union[int, float]
:param c: Variable to control parameter adoption. Must be in [0, 1].
:type c: Union[int, float]
:type seed: Union[int, None]
:return: A pair with the best solution found and its fitness.
:rtype [np.ndarray, int]
"""
# 0. Check external parameters
if type(population_size) is not int or population_size <= 0:
raise ValueError("population_size must be a positive integer.")
if type(individual_size) is not int or individual_size <= 0:
raise ValueError("individual_size must be a positive integer.")
if type(max_evals) is not int or max_evals <= 0:
raise ValueError("max_evals must be a positive integer.")
if type(bounds) is not np.ndarray or bounds.shape != (individual_size, 2):
raise ValueError("bounds must be a NumPy ndarray.\n"
"The array must be of individual_size length. "
"Each row must have 2 elements.")
if type(seed) is not int and seed is not None:
raise ValueError("seed must be an integer or None.")
if type(p) not in [int, float] and 0 < p <= 1:
raise ValueError("p must be a real number in (0, 1].")
if type(c) not in [int, float] and 0 <= c <= 1:
raise ValueError("c must be an real number in [0, 1].")
if type(ng) is not int:
raise ValueError("ng must be a positive integer number.")
if type(lambdas) not in [list, np.ndarray
] and len(lambdas) != 4 and sum(lambdas) != 1:
raise ValueError(
"lambdas must be a list or npdarray of 4 numbers that sum 1.")
np.random.seed(seed)
# 1. Initialize internal parameters
# 1.1 Control parameters
u_cr = np.ones(3) * 0.5
u_f = np.ones(3) * 0.5
f_var = np.zeros(3)
fes = np.zeros(3)
# 1.2 Initialize population
pop_size = lambdas * population_size
if population is None:
big_population = commons.init_population(int(sum(pop_size)),
individual_size, bounds)
pops = np.array_split(big_population, 4)
chosen = np.random.randint(0, 3)
newpop = np.concatenate((pops[chosen], pops[3]))
pops[chosen] = newpop
pop_size = list(map(len, pops))
current_generation = 0
num_evals = 0
f = []
cr = []
fitnesses = []
for j in range(3):
f.append(np.empty(pop_size[j]))
cr.append(np.empty(pop_size[j]))
fitnesses.append(commons.apply_fitness(pops[j], func, opts))
num_evals += len(pops[j])
# 2. Start the algorithm
while num_evals <= max_evals:
current_generation += 1
# 2.1 Generate CR and F values
for j in range(3):
f[j] = scipy.stats.cauchy.rvs(loc=u_f[j],
scale=0.1,
size=len(pops[j]))
f[j] = np.clip(f[j], 0, 1)
cr[j] = np.random.normal(u_cr[j], 0.1, len(pops[j]))
cr[j] = np.clip(cr[j], 0, 1)
# 2.2 Apply mutation to each subpopulation
mutated1 = commons.current_to_pbest_mutation(
pops[0], fitnesses[0], f[0].reshape(len(f[0]), 1),
np.ones(len(pops[0])) * p, bounds)
mutated2 = commons.current_to_rand_1_mutation(
pops[1], fitnesses[1], f[1].copy().reshape(len(f[1]), 1) * .5 + 1,
f[1].reshape(len(f[1]), 1), bounds)
mutated3 = commons.binary_mutation(pops[2], f[2].reshape(len(f[2]), 1),
bounds)
# 2.3 Do the crossover and calculate new fitness
crossed1 = commons.crossover(pops[0], mutated1,
cr[0].reshape(len(cr[0]), 1))
crossed2 = mutated2
crossed3 = commons.crossover(pops[2], mutated3,
cr[2].reshape(len(cr[2]), 1))
c_fitness1 = commons.apply_fitness(crossed1, func, opts)
c_fitness2 = commons.apply_fitness(crossed2, func, opts)
c_fitness3 = commons.apply_fitness(crossed3, func, opts)
for j in range(3):
num_evals += len(pops[j])
fes[j] += len(pops[j])
# 2.4 Do the selection and update control parameters
winners1 = c_fitness1 < fitnesses[0]
winners2 = c_fitness2 < fitnesses[1]
winners3 = c_fitness3 < fitnesses[2]
pops[0] = commons.selection(pops[0], crossed1, fitnesses[0],
c_fitness1)
pops[1] = commons.selection(pops[1], crossed2, fitnesses[1],
c_fitness2)
pops[2] = commons.selection(pops[2], crossed3, fitnesses[2],
c_fitness3)
fitnesses[0][winners1] = c_fitness1[winners1]
fitnesses[1][winners2] = c_fitness2[winners2]
fitnesses[2][winners3] = c_fitness3[winners3]
if sum(winners1) != 0 and np.sum(f[0][winners1]) != 0:
u_cr[0] = (1 - c) * u_cr[0] + c * np.mean(cr[0][winners1])
u_f[0] = (1 - c) * u_f[0] + c * (np.sum(f[0][winners1]**2) /
np.sum(f[0][winners1]))
if sum(winners2) != 0 and np.sum(f[1][winners2]) != 0:
u_cr[1] = (1 - c) * u_cr[1] + c * np.mean(cr[1][winners2])
u_f[1] = (1 - c) * u_f[1] + c * (np.sum(f[1][winners2]**2) /
np.sum(f[1][winners2]))
if sum(winners3) != 0 and np.sum(f[2][winners3]) != 0:
u_cr[2] = (1 - c) * u_cr[2] + c * np.mean(cr[2][winners3])
u_f[2] = (1 - c) * u_f[2] + c * (np.sum(f[2][winners3]**2) /
np.sum(f[2][winners3]))
fes[0] += np.sum(fitnesses[0][winners1] - c_fitness1[winners1])
fes[1] += np.sum(fitnesses[1][winners2] - c_fitness2[winners2])
fes[2] += np.sum(fitnesses[2][winners3] - c_fitness3[winners3])
population = np.concatenate((pops[0], pops[1], pops[2]))
fitness = np.concatenate((fitnesses[0], fitnesses[1], fitnesses[2]))
if current_generation % ng == 0:
k = [f_var[i] / len(pops[i] / ng) for i in range(3)]
chosen = np.argmax(k)
indexes = np.arange(0, len(population), 1, np.int)
np.random.shuffle(indexes)
indexes = np.array_split(indexes, 4)
chosen = np.random.randint(0, 3)
pops = []
fitnesses = []
f = []
cr = []
for j in range(3):
if j == chosen:
pops.append(
np.concatenate(
(population[indexes[j]], population[indexes[3]])))
fitnesses.append(
np.concatenate((fitness[indexes[j]], fitness[indexes[3]])))
else:
pops.append(population[indexes[j]])
fitnesses.append(fitness[indexes[j]])
f.append(np.empty(len(pops[j])))
cr.append(np.empty(len(pops[j])))
if callback is not None:
callback(**(locals()))
best = np.argmin(fitness)
if fitness[best] == answer:
yield population[best], fitness[best], population, fitness
break
else:
yield population[best], fitness[best], population, fitness
def apply(population_size: int, individual_size: int, bounds: np.ndarray,
func: Callable[[np.ndarray], np.float], opts: Any, memory_size: int,
callback: Callable[[Dict], Any], max_evals: int, seed: Union[int,
None],
population: Union[np.array,
None], answer: Union[None, float,
int]) -> [np.ndarray, int]:
"""
Applies the L-SHADE Differential Evolution Algorithm.
:param population_size: Size of the population.
:type population_size: int
:param individual_size: Number of gens/features of an individual.
:type individual_size: int
:param bounds: Numpy ndarray with individual_size rows and 2 columns.
First column represents the minimum value for the row feature.
Second column represent the maximum value for the row feature.
:type bounds: np.ndarray
:param func: Evaluation function. The function used must receive one
parameter.This parameter will be a numpy array representing an individual.
:type func: Callable[[np.ndarray], float]
:param opts: Optional parameters for the fitness function.
:type opts: Any type.
:param memory_size: Size of the internal memory.
:type memory_size: int
:param callback: Optional function that allows read access to the state of all variables once each generation.
:type callback: Callable[[Dict], Any]
:param max_evals: Number of evaluations after the algorithm is stopped.
:type max_evals: int
:param seed: Random number generation seed. Fix a number to reproduce the
same results in later experiments.
:type seed: Union[int, None]
:return: A pair with the best solution found and its fitness.
:rtype [np.ndarray, int]
"""
# 0. Check parameters are valid
if type(population_size) is not int or population_size <= 0:
raise ValueError("population_size must be a positive integer.")
if type(individual_size) is not int or individual_size <= 0:
raise ValueError("individual_size must be a positive integer.")
if type(max_evals) is not int or max_evals <= 0:
raise ValueError("max_iter must be a positive integer.")
if type(bounds) is not np.ndarray or bounds.shape != (individual_size, 2):
raise ValueError("bounds must be a NumPy ndarray.\n"
"The array must be of individual_size length. "
"Each row must have 2 elements.")
if type(seed) is not int and seed is not None:
raise ValueError("seed must be an integer or None.")
np.random.seed(seed)
random.seed(seed)
# 1. Initialization
if population is None:
population = commons.init_population(population_size, individual_size,
bounds)
init_size = population_size
m_cr = np.ones(memory_size) * 0.5
m_f = np.ones(memory_size) * 0.5
archive = []
k = 0
fitness = commons.apply_fitness(population, func, opts)
all_indexes = list(range(memory_size))
current_generation = 0
num_evals = population_size
# Calculate max_iters
n = population_size
i = 0
max_iters = 0
# Calculate max_iters
n = population_size
i = 0
max_iters = 0
while i < max_evals:
max_iters += 1
n = round((4 - init_size) / max_evals * i + init_size)
i += n
while num_evals < max_evals:
# 2.1 Adaptation
r = np.random.choice(all_indexes, population_size)
cr = np.random.normal(m_cr[r], 0.1, population_size)
cr = np.clip(cr, 0, 1)
cr[m_cr[r] == 1] = 0
f = scipy.stats.cauchy.rvs(loc=m_f[r], scale=0.1, size=population_size)
f[f > 1] = 0
while sum(f <= 0) != 0:
r = np.random.choice(all_indexes, sum(f <= 0))
f[f <= 0] = scipy.stats.cauchy.rvs(loc=m_f[r],
scale=0.1,
size=sum(f <= 0))
p = np.ones(population_size) * .11
# 2.2 Common steps
mutated = commons.current_to_pbest_mutation(population, fitness,
f.reshape(len(f), 1), p,
bounds)
crossed = commons.crossover(population, mutated, cr.reshape(len(f), 1))
c_fitness = commons.apply_fitness(crossed, func, opts)
num_evals += population_size
population, indexes = commons.selection(population,
crossed,
fitness,
c_fitness,
return_indexes=True)
# 2.3 Adapt for next generation
archive.extend(population[indexes])
if len(indexes) > 0:
if len(archive) > population_size:
archive = random.sample(archive, population_size)
weights = np.abs(fitness[indexes] - c_fitness[indexes])
weights /= np.sum(weights)
m_cr[k] = np.sum(weights * cr[indexes]**2) / np.sum(
weights * cr[indexes])
if np.isnan(m_cr[k]):
m_cr[k] = 1
m_f[k] = np.sum(weights * f[indexes]**2) / np.sum(
weights * f[indexes])
k += 1
if k == memory_size:
k = 0
fitness[indexes] = c_fitness[indexes]
# Adapt population size
new_population_size = round((4 - init_size) / max_evals * num_evals +
init_size)
if population_size > new_population_size:
population_size = new_population_size
best_indexes = np.argsort(fitness)[:population_size]
population = population[best_indexes]
fitness = fitness[best_indexes]
if k == init_size:
k = 0
if callback is not None:
callback(**(locals()))
current_generation += 1
best = np.argmin(fitness)
if fitness[best] == answer:
yield population[best], fitness[best], population, fitness
break
else:
yield population[best], fitness[best], population, fitness
def apply(population_size: int, individual_size: int, bounds: np.ndarray,
func: Callable[[np.ndarray], float], opts: Any, p: Union[int, float],
c: Union[int, float], callback: Callable[[Dict], Any],
max_evals: int, seed: Union[int, None], population: Union[np.array,
None],
answer: Union[None, int, float]) -> [np.ndarray, int]:
"""
Applies the JADE Differential Evolution algorithm.
:param population_size: Size of the population.
:type population_size: int
:param individual_size: Number of gens/features of an individual.
:type individual_size: int
:param bounds: Numpy ndarray with individual_size rows and 2 columns.
First column represents the minimum value for the row feature.
Second column represent the maximum value for the row feature.
:type bounds: np.ndarray
:param func: Evaluation function. The function used must receive one
parameter.This parameter will be a numpy array representing an individual.
:type func: Callable[[np.ndarray], float]
:param opts: Optional parameters for the fitness function.
:type opts: Any type.
:param p: Parameter to choose the best vectors. Must be in (0, 1].
:type p: Union[int, float]
:param c: Variable to control parameter adoption. Must be in [0, 1].
:type c: Union[int, float]
:param callback: Optional function that allows read access to the state of all variables once each generation.
:type callback: Callable[[Dict], Any]
:param max_evals: Number of evaluations after the algorithm is stopped.
:type max_evals: int
:param seed: Random number generation seed. Fix a number to reproduce the
same results in later experiments.
:type seed: Union[int, None]
:return: A pair with the best solution found and its fitness.
:rtype [np.ndarray, int]
"""
# 0. Check parameters are valid
if type(population_size) is not int or population_size <= 0:
raise ValueError("population_size must be a positive integer.")
if type(individual_size) is not int or individual_size <= 0:
raise ValueError("individual_size must be a positive integer.")
if type(max_evals) is not int or max_evals <= 0:
raise ValueError("max_evals must be a positive integer.")
if type(bounds) is not np.ndarray or bounds.shape != (individual_size, 2):
raise ValueError("bounds must be a NumPy ndarray.\n"
"The array must be of individual_size length. "
"Each row must have 2 elements.")
if type(seed) is not int and seed is not None:
raise ValueError("seed must be an integer or None.")
if type(p) not in [int, float] and 0 < p <= 1:
raise ValueError("p must be a real number in (0, 1].")
if type(c) not in [int, float] and 0 <= c <= 1:
raise ValueError("c must be an real number in [0, 1].")
np.random.seed(seed)
# 1. Init population
if population is None:
population = commons.init_population(population_size, individual_size,
bounds)
u_cr = 0.5
u_f = 0.6
p = np.ones(population_size) * p
fitness = commons.apply_fitness(population, func, opts)
max_iters = max_evals // population_size
#print("jade : ", max_iters)
results = []
for current_generation in range(max_iters):
# 2.1 Generate parameter values for current generation
cr = np.random.normal(u_cr, 0.1, population_size)
f = np.random.rand(population_size // 3) * 1.2
f = np.concatenate(
(f,
np.random.normal(u_f, 0.1,
population_size - (population_size // 3))))
# 2.2 Common steps
mutated = commons.current_to_pbest_mutation(population, fitness,
f.reshape(len(f), 1), p,
bounds)
crossed = commons.crossover(population, mutated, cr.reshape(len(f), 1))
c_fitness = commons.apply_fitness(crossed, func, opts)
population, indexes = commons.selection(population,
crossed,
fitness,
c_fitness,
return_indexes=True)
# 2.3 Adapt for next generation
if len(indexes) != 0:
u_cr = (1 - c) * u_cr + c * np.mean(cr[indexes])
u_f = (1 - c) * u_f + c * (np.sum(f[indexes]**2) /
np.sum(f[indexes]))
fitness[indexes] = c_fitness[indexes]
if callback is not None:
callback(**(locals()))
best = np.argmin(fitness)
if fitness[best] == answer:
results.append(
(population[best], fitness[best], population, fitness))
break
else:
results.append(
(population[best], fitness[best], population, fitness))
return results
def apply(population_size: int, individual_size: int, f: Union[float, int],
cr: Union[float,
int], bounds: np.ndarray, func: Callable[[np.ndarray],
float], opts: Any,
callback: Callable[[Dict], Any], cross: str, max_evals: int,
seed: Union[int, None], population: Union[np.ndarray, None],
answer: Union[None, int, float]) -> [np.ndarray, int]:
"""
Applies the standard differential evolution algorithm.
:param population_size: Size of the population.
:type population_size: int
:param individual_size: Number of gens/features of an individual.
:type individual_size: int
:param f: Mutation parameter. Must be in [0, 2].
:type f: Union[float, int]
:param cr: Crossover Ratio. Must be in [0, 1].
:type cr: Union[float, int]
:param bounds: Numpy ndarray with individual_size rows and 2 columns.
First column represents the minimum value for the row feature.
Second column represent the maximum value for the row feature.
:type bounds: np.ndarray
:param func: Evaluation function. The function used must receive one
parameter.This parameter will be a numpy array representing an individual.
:type func: Callable[[np.ndarray], float]
:param opts: Optional parameters for the fitness function.
:type opts: Any type.
:param callback: Optional function that allows read access to the state of all variables once each generation.
:type callback: Callable[[Dict], Any]
:param cross: Indicates whether to use the binary crossover('bin') or the exponential crossover('exp').
:type cross: str
:param max_evals: Number of evaluations after the algorithm is stopped.
:type max_evals: int
:param seed: Random number generation seed. Fix a number to reproduce the
same results in later experiments.
:type seed: Union[int, None]
:return: A pair with the best solution found and its fitness.
:rtype [np.ndarray, int]
"""
# 0. Check parameters are valid
if type(population_size) is not int or population_size <= 0:
raise ValueError("population_size must be a positive integer.")
if type(individual_size) is not int or individual_size <= 0:
raise ValueError("individual_size must be a positive integer.")
if (type(f) is not int and type(f) is not float) or not 0 <= f <= 2:
raise ValueError("f (mutation parameter) must be a "
"real number in [0,2].")
if (type(cr) is not int and type(cr) is not float) or not 0 <= cr <= 1:
raise ValueError("cr (crossover ratio) must be a "
"real number in [0,1].")
if type(max_evals) is not int or max_evals <= 0:
raise ValueError("max_evals must be a positive integer.")
if type(bounds) is not np.ndarray or bounds.shape != (individual_size, 2):
raise ValueError("bounds must be a NumPy ndarray.\n"
"The array must be of individual_size length. "
"Each row must have 2 elements.")
if type(cross) is not str and cross not in ['bin', 'exp']:
raise ValueError(
"cross must be a string and must be one of \'bin\' or \'cross\'")
if type(seed) is not int and seed is not None:
raise ValueError("seed must be an integer or None.")
# 1. Initialization
np.random.seed(seed)
if population is None:
population = commons.init_population(population_size, individual_size,
bounds)
#population = commons.init_population(population_size, individual_size, bounds)
try:
fitness = commons.apply_fitness(population, func, opts)
except TypeError:
print(func, population)
#use self.population and self.fitness - move the loop into a step function
max_iters = max_evals // population_size
#print("de : ", max_iters)
results = []
for current_generation in range(max_iters):
mutated = commons.binary_mutation(population, f, bounds)
if cross == 'bin':
crossed = commons.crossover(population, mutated, cr)
else:
crossed = commons.exponential_crossover(population, mutated, cr)
c_fitness = commons.apply_fitness(crossed, func, opts)
population, indexes = commons.selection(population,
crossed,
fitness,
c_fitness,
return_indexes=True)
fitness[indexes] = c_fitness[indexes]
#print(locals())
best = np.argmin(fitness)
if callback is not None:
callback(**(locals()))
best = np.argmin(fitness)
if fitness[best] == answer:
results.append(
(population[best], fitness[best], population, fitness))
break
else:
results.append(
(population[best], fitness[best], population, fitness))
return results
def apply(population_size: int, individual_size: int, bounds: np.ndarray,
func: Callable[[np.ndarray], float], opts: Any,
callback: Callable[[Dict], Any], max_evals: int, seed: Union[int,
None],
answer: Union[None, int, float], population: [np.ndarray, None]):
"""
Applies the Self-adaptive differential evolution algorithm (SaDE).
:param population_size: Size of the population.
:type population_size: int
:param individual_size: Number of gens/features of an individual.
:type individual_size: int
:param bounds: Numpy ndarray with individual_size rows and 2 columns.
First column represents the minimum value for the row feature.
Second column represent the maximum value for the row feature.
:type bounds: np.ndarray
:param func: Evaluation function. The function used must receive one
parameter.This parameter will be a numpy array representing an individual.
:type func: Callable[[np.ndarray], float]
:param opts: Optional parameters for the fitness function.
:type opts: Any type.
:param callback: Optional function that allows read access to the state of all variables once each generation.
:type callback: Callable[[Dict], Any]
:param max_evals: Number of evaluatios after the algorithm is stopped.
:type max_evals: int
:param seed: Random number generation seed. Fix a number to reproduce the
same results in later experiments.
:type seed: Union[int, None]
:return: A pair with the best solution found and its fitness.
:rtype [np.ndarray, int]
"""
if type(population_size) is not int or population_size <= 0:
raise ValueError("population_size must be a positive integer.")
if type(individual_size) is not int or individual_size <= 0:
raise ValueError("individual_size must be a positive integer.")
if type(bounds) is not np.ndarray or bounds.shape != (individual_size, 2):
raise ValueError("bounds must be a NumPy ndarray.\n"
"The array must be of individual_size length. "
"Each row must have 2 elements.")
if type(max_evals) is not int or max_evals <= 0:
raise ValueError("max_evals must be a positive integer.")
if type(seed) is not int and seed is not None:
raise ValueError("seed must be an integer or None.")
# 1. Initialization
np.random.seed(seed)
if population is None:
population = commons.init_population(population_size, individual_size,
bounds)
# 2. SaDE Algorithm
probability = 0.5
fitness = commons.apply_fitness(population, func, opts)
cr_m = 0.5
f_m = 0.5
sum_ns1 = 0
sum_nf1 = 0
sum_ns2 = 0
sum_nf2 = 0
cr_list = []
f = np.random.normal(f_m, 0.3, population_size)
f = np.clip(f, 0, 2)
cr = np.random.normal(cr_m, 0.1, population_size)
cr = np.clip(cr, 0, 1)
max_iters = max_evals // population_size
#print("sade : ", max_iters)
results = []
for current_generation in range(max_iters):
# 2.1 Mutation
# 2.1.1 Randomly choose which individuals do each mutation
choice = np.random.rand(population_size)
choice_1 = choice < probability
choice_2 = choice >= probability
# 2.1.2 Apply the mutations
mutated = population.copy()
mutated[choice_1] = commons.binary_mutation(
population[choice_1], f[choice_1].reshape(sum(choice_1), 1),
bounds)
mutated[choice_2] = commons.current_to_best_2_binary_mutation(
population[choice_2], fitness[choice_2],
f[choice_2].reshape(sum(choice_2), 1), bounds)
# 2.2 Crossover
crossed = commons.crossover(population, mutated,
cr.reshape(population_size, 1))
c_fitness = commons.apply_fitness(crossed, func, opts)
# 2.3 Selection
population = commons.selection(population, crossed, fitness, c_fitness)
winners = c_fitness < fitness
fitness[winners] = c_fitness[winners]
# 2.4 Self Adaption
chosen_1 = np.sum(np.bitwise_and(choice_1, winners))
chosen_2 = np.sum(np.bitwise_and(choice_2, winners))
sum_ns1 += chosen_1
sum_ns2 += chosen_2
sum_nf1 += np.sum(choice_1) - chosen_1
sum_nf2 += np.sum(choice_2) - chosen_2
cr_list = np.concatenate((cr_list, cr[winners]))
# 2.4.1 Adapt mutation strategy probability
if (current_generation + 1) % 50 == 0:
probability = sum_ns1 * (sum_ns2 + sum_nf2) / (sum_ns2 *
(sum_ns1 + sum_nf1))
probability = np.clip(probability, 0, 1)
if np.isnan(probability):
probability = .99
sum_ns1 = 0
sum_ns2 = 0
sum_nf1 = 0
sum_nf2 = 0
# 2.4.2
if (current_generation + 1) % 25 == 0:
if len(cr_list) != 0:
cr_m = np.mean(cr_list)
cr_list = []
cr = np.random.normal(cr_m, 0.1, population_size)
cr = np.clip(cr, 0, 1)
if callback is not None:
callback(**(locals()))
best = np.argmin(fitness)
if fitness[best] == answer:
results.append(
(population[best], fitness[best], population, fitness))
break
else:
results.append(
(population[best], fitness[best], population, fitness))
return results
