def genetic_algorithm(instance,
config,
fitness_function,
*,
best_fitness=None,
perf_counter=None,
process_time=None,
all_fitness=None):
population_size = config.population_size
population = np.random.randint(2,
size=(population_size,
instance.num_materials),
dtype=bool)
# selection_quant = 2
# if (config.crossover_method == Crossover.THREE_PARENT_CROSSOVER):
# selection_quant = 3
timer = Timer(
) ########################################################################################################
start_perf_counter = time.perf_counter()
start_process_time = time.process_time()
for iteration in range(config.num_iterations):
timer.add_time(
) ########################################################################################################
# print('==========================' + str(iteration))
survival_values = np.apply_along_axis(fitness_function,
1,
population,
instance,
timer,
data=all_fitness)
sorted_indices = np.argsort(survival_values)
population = population[sorted_indices]
survival_values = survival_values[sorted_indices]
if best_fitness is not None:
best_fitness[iteration] = survival_values[0]
# best_fitness[iteration] = np.mean(survival_values)
# best_fitness[iteration] = survival_values[-1]
if perf_counter is not None:
perf_counter[iteration] = time.perf_counter() - start_perf_counter
if process_time is not None:
process_time[iteration] = time.process_time() - start_process_time
new_population = copying_gene(population, config.copying_method,
config)
if config.use_local_search:
new_population = local_search_gene(new_population,
fitness_function,
config.local_search_method,
config)
remaining_spots = np.random.randint(
2,
size=(population_size - new_population.shape[0],
instance.num_materials),
dtype=bool)
remaining_spots = population_size - len(new_population)
selection_spots = remaining_spots
if (config.crossover_method == Crossover.THREE_PARENT_CROSSOVER):
selection_spots = int(3 * math.ceil(remaining_spots / 3.)) * 3
elif (config.crossover_method == Crossover.UNIFORM_CROSSOVER):
selection_spots = int(2 * math.ceil(remaining_spots / 2.)) * 2
else:
selection_spots = int(2 * math.ceil(remaining_spots / 2.))
parents = selection_gene(population, survival_values, selection_spots,
config.selection_method, config)
# children = crossover_gene(parents, remaining_spots, config.crossover_method, config)
children = crossover_gene(parents, config.crossover_method, config)
assert parents.shape[0] == remaining_spots
mutated = mutation_gene(children, config.mutation_method, config)
np.append(new_population, mutated, axis=0)
return (0, 0)
def particle_swarm_optmization(instance, config, fitness_function, out_info=None):
num_particles = config.num_particles
if config.evaluator == Evaluator.FIXED_EVALUATOR:
evaluate_function = evaluate_population_fixed
else:
evaluate_function = evaluate_population_random
def counter_fitness(individual, instance, student, timer, print_results=False, data=None):
nonlocal cost_counter
cost_counter += 1
return fitness_function(individual, instance, student, timer, print_results, data=data)
if out_info is not None:
out_info['best_fitness'] = []
out_info['partial_fitness'] = []
out_info['perf_counter'] = []
out_info['process_time'] = []
out_info['cost_value'] = []
results = []
for student in range(instance.num_learners):
cost_counter = 0
iteration_counter = 0
stagnation_counter = 0
if out_info is not None:
out_info['best_fitness'].append([])
out_info['partial_fitness'].append([])
out_info['perf_counter'].append([])
out_info['process_time'].append([])
out_info['cost_value'].append([])
timer = Timer()
timer.add_time()
particle_velocity = np.random.rand(num_particles, instance.num_materials) * (2 * config.max_velocity) - config.max_velocity
particle_position = evaluate_function(particle_velocity)
local_best_position = np.copy(particle_position)
local_best_fitness = np.apply_along_axis(counter_fitness, 1, local_best_position, instance, student, timer)
global_best_index = np.argmin(local_best_fitness, axis=0)
global_best_position = np.copy(local_best_position[global_best_index])
global_best_fitness = local_best_fitness[global_best_index]
timer.add_time("initialization")
start_perf_counter = time.perf_counter()
start_process_time = time.process_time()
while ((not config.cost_budget or cost_counter < config.cost_budget) and
(not config.num_iterations or iteration_counter < config.num_iterations) and
(not config.max_stagnation or stagnation_counter < config.max_stagnation)):
old_global_best_fitness = global_best_fitness
if out_info is not None:
out_info["best_fitness"][-1].append(global_best_fitness)
fitness_function(global_best_position, instance, student, timer, data=out_info["partial_fitness"][-1])
out_info["perf_counter"][-1].append(time.perf_counter() - start_perf_counter)
out_info["process_time"][-1].append(time.process_time() - start_process_time)
out_info["cost_value"][-1].append(cost_counter)
timer.add_time()
local_influence = np.tile(config.local_influence_parameter * np.random.random(num_particles), (instance.num_materials, 1)).T
global_influence = np.tile(config.global_influence_parameter * np.random.random(num_particles), (instance.num_materials, 1)).T
local_distance = local_best_position.astype(int) - particle_position.astype(int)
global_distance = global_best_position.astype(int) - particle_position.astype(int)
particle_velocity = (config.inertia_parameter * particle_velocity
+ local_influence * local_distance
+ global_influence * global_distance)
particle_velocity = np.clip(particle_velocity, -config.max_velocity, config.max_velocity)
timer.add_time("update_velocity")
# Calcula as novas posições
particle_position = evaluate_function(particle_velocity)
timer.add_time("update_position")
# Calcula os novos resultados
particle_new_fitness = np.apply_along_axis(counter_fitness, 1, particle_position, instance, student, timer)
timer.add_time("update_fitness")
# Calcula a mascara de melhores valores para cada particula
change_mask = (particle_new_fitness < local_best_fitness)
# Altera o melhor resultado de cada particula
local_best_position[change_mask] = np.copy(particle_position[change_mask])
local_best_fitness[change_mask] = particle_new_fitness[change_mask]
# Encontra o melhor resultado entre todas as particulas
global_best_index = np.argmin(local_best_fitness)
global_best_position = np.copy(local_best_position[global_best_index])
global_best_fitness = local_best_fitness[global_best_index]
iteration_counter += 1
if global_best_fitness < old_global_best_fitness:
stagnation_counter = 0
else:
stagnation_counter += 1
timer.add_time("update_best")
if out_info is not None:
out_info["best_fitness"][-1].append(global_best_fitness)
fitness_function(global_best_position, instance, student, timer, data=out_info["partial_fitness"][-1])
out_info["perf_counter"][-1].append(time.perf_counter() - start_perf_counter)
out_info["process_time"][-1].append(time.process_time() - start_process_time)
out_info["cost_value"][-1].append(cost_counter)
results.append((global_best_position, global_best_fitness))
return results
def prey_predator_algorithm_discrete(instance, config, fitness_function, out_info=None):
population_size = config.population_size
if config.max_steps > instance.num_materials:
config.max_steps = instance.num_materials
if config.min_steps > config.max_steps:
config.min_steps = config.max_steps
def counter_fitness(individual, instance, student, timer, print_results=False, data=None):
nonlocal cost_counter
cost_counter += 1
return fitness_function(individual, instance, student, timer, print_results, data=data)
if out_info is not None:
out_info["best_fitness"] = []
out_info["partial_fitness"] = []
out_info["perf_counter"] = []
out_info["process_time"] = []
out_info["cost_value"] = []
results = []
for student in range(instance.num_learners):
cost_counter = 0
iteration_counter = 0
stagnation_counter = 0
if out_info is not None:
out_info['best_fitness'].append([])
out_info['partial_fitness'].append([])
out_info['perf_counter'].append([])
out_info['process_time'].append([])
out_info['cost_value'].append([])
timer = Timer()
population = np.random.randint(2, size=(population_size, instance.num_materials), dtype=bool)
population_best_individual = population[0]
population_best_fitness = counter_fitness(population[0], instance, student, timer)
start_perf_counter = time.perf_counter()
start_process_time = time.process_time()
while ((not config.cost_budget or cost_counter < config.cost_budget) and
(not config.num_iterations or iteration_counter < config.num_iterations) and
(not config.max_stagnation or stagnation_counter < config.max_stagnation)):
timer.add_time()
survival_values = np.apply_along_axis(counter_fitness, 1, population, instance, student, timer)
sorted_indices = np.argsort(survival_values)
population = population[sorted_indices]
survival_values = survival_values[sorted_indices]
iteration_counter += 1
if survival_values[0] < population_best_fitness:
population_best_individual = population[0]
population_best_fitness = survival_values[0]
stagnation_counter = 0
else:
stagnation_counter += 1
if out_info is not None:
out_info["best_fitness"][-1].append(population_best_fitness)
fitness_function(population_best_individual, instance, student, timer, data=out_info["partial_fitness"][-1])
out_info["perf_counter"][-1].append(time.perf_counter() - start_perf_counter)
out_info["process_time"][-1].append(time.process_time() - start_process_time)
out_info["cost_value"][-1].append(cost_counter)
new_population = np.copy(population)
timer.add_time("creation")
# Cria as mascaras para separar os diferentes tipos de individuo
best_prey_mask = np.zeros(population_size, dtype=bool)
best_prey_mask[0] = True
predator_mask = np.zeros(population_size, dtype=bool)
predator_mask[-1] = True
follow_mask = (np.random.rand(population_size) < config.follow_chance)
run_mask = ~follow_mask
follow_mask[best_prey_mask] = False # Ignora as melhores presas
follow_mask[predator_mask] = False # Ignora os predadores
run_mask[best_prey_mask] = False # Ignora as melhores presas
run_mask[predator_mask] = False # Ignora os predadores
timer.add_time()
follow_indices = np.where(follow_mask)[0]
follow_quant = len(follow_indices)
population_distance = [[hamming_distance(population[j], population[i]) / instance.num_materials for j in range(i)] for i in follow_indices]
survival_ratio = [[survival_values[j] / survival_values[i] for j in range(i)] for i in follow_indices]
follow_chance = [[(2 - config.follow_distance_parameter * population_distance[i][j] - config.follow_survival_parameter * survival_ratio[i][j]) / 2 for j in range(follow_indices[i])] for i in range(follow_quant)]
roulette_array = np.array([Roulette(follow_chance[i]) for i in range(follow_quant)])
timer.add_time("follow_chance")
# TODO(andre:2018-05-28): Garantir que max_steps nunca é maior do que o numero de materiais
# TODO(andre:2018-12-20): Verificar o calculo do número de passos. Ele está usando a distância até a proxima presa e não a distância até o predador
num_steps = np.round(config.max_steps * np.random.rand(follow_quant) / np.exp(config.steps_distance_parameter * np.array([i[-1] for i in population_distance])))
new_population[follow_mask] = move_population_roulette(new_population[follow_mask], num_steps, roulette_array, population)
timer.add_time("follow_roulette")
num_steps = np.round(config.min_steps * np.random.rand(follow_quant))
new_population[follow_mask] = move_population_random(new_population[follow_mask], num_steps)
timer.add_time("follow_random")
num_steps = np.round(config.max_steps * np.random.rand(np.count_nonzero(run_mask)))
new_population[run_mask] = move_population_random_complement(new_population[run_mask], num_steps, population[-1])
timer.add_time("run")
new_population[best_prey_mask] = move_population_local_search(new_population[best_prey_mask], counter_fitness, config.min_steps, config.local_search_tries, instance, student, timer)
timer.add_time()
num_steps = np.round(config.max_steps * np.random.rand(np.count_nonzero(predator_mask)))
new_population[predator_mask] = move_population_random(new_population[predator_mask], num_steps)
timer.add_time("predator_random")
num_steps = np.round(config.min_steps * np.random.rand(np.count_nonzero(predator_mask)))
worst_prey = np.repeat(population[-2][np.newaxis, :], np.count_nonzero(predator_mask), axis=0)
new_population[predator_mask] = move_population_direction(new_population[predator_mask], num_steps, worst_prey)
timer.add_time("predator_follow")
population = new_population
survival_values = np.apply_along_axis(counter_fitness, 1, population, instance, student, timer)
sorted_indices = np.argsort(survival_values)
population = population[sorted_indices]
survival_values = survival_values[sorted_indices]
if out_info is not None:
out_info["best_fitness"][-1].append(population_best_fitness)
fitness_function(population_best_individual, instance, student, timer, data=out_info["partial_fitness"][-1])
out_info["perf_counter"][-1].append(time.perf_counter() - start_perf_counter)
out_info["process_time"][-1].append(time.process_time() - start_process_time)
out_info["cost_value"][-1].append(cost_counter)
results.append((population_best_individual, population_best_fitness))
return results
def particle_swarm_optmization(instance, config, fitness_function, *, best_fitness=None, perf_counter=None, process_time=None, all_fitness=None):
num_particles = config.num_particles
timer = Timer() ########################################################################################################
timer.add_time() ########################################################################################################
particle_velocity = np.random.rand(num_particles, instance.num_materials) * (2 * config.max_velocity) - config.max_velocity
particle_position = generate_particle_position(particle_velocity)
local_best_position = np.copy(particle_position)
local_best_fitness = np.apply_along_axis(fitness_function, 1, local_best_position, instance, timer, data=all_fitness)
global_best_index = np.argmin(local_best_fitness, axis=0)
global_best_position = np.copy(local_best_position[global_best_index])
global_best_fitness = local_best_fitness[global_best_index]
timer.add_time("initialization") ########################################################################################################
start_perf_counter = time.perf_counter()
start_process_time = time.process_time()
for iteration in range(config.num_iterations):
if best_fitness is not None:
best_fitness[iteration] = global_best_fitness
# best_fitness[iteration] = np.mean(survival_values)
# best_fitness[iteration] = survival_values[-1]
if perf_counter is not None:
perf_counter[iteration] = time.perf_counter() - start_perf_counter
if process_time is not None:
process_time[iteration] = time.process_time() - start_process_time
timer.add_time() ########################################################################################################
# TODO(andre:2018-04-18): Atualizar velocidade
local_influence = np.tile(config.local_influence_parameter * np.random.random(num_particles), (instance.num_materials, 1)).T
global_influence = np.tile(config.global_influence_parameter * np.random.random(num_particles), (instance.num_materials, 1)).T
local_distance = local_best_position.astype(int) - particle_position.astype(int)
global_distance = global_best_position.astype(int) - particle_position.astype(int)
# print(particle_velocity[0])
particle_velocity = (config.inertia_parameter * particle_velocity
+ local_influence * local_distance
+ global_influence * global_distance)
particle_velocity = np.clip(particle_velocity, -config.max_velocity, config.max_velocity)
timer.add_time("update_velocity") ########################################################################################################
# Calcula as novas posições
particle_position = generate_particle_position(particle_velocity)
timer.add_time("update_position") ########################################################################################################
# Calcula os novos resultados
particle_new_fitness = np.apply_along_axis(fitness, 1, particle_position, instance, timer, data=all_fitness)
timer.add_time("update_fitness") ########################################################################################################
# Calcula a mascara de melhores valores para cada particula
change_mask = (particle_new_fitness < local_best_fitness)
# Altera o melhor resultado de cada particula
local_best_position[change_mask] = np.copy(particle_position[change_mask])
# local_best_position[change_mask] = particle_position[change_mask]
local_best_fitness[change_mask] = particle_new_fitness[change_mask]
# Encontra o melhor resultado entre todas as particulas
global_best_index = np.argmin(local_best_fitness)
global_best_position = np.copy(local_best_position[global_best_index])
global_best_fitness = local_best_fitness[global_best_index]
timer.add_time("update_best") ########################################################################################################
print("Tempo: ")
print(timer.get_time())
print("Iterações: ")
print(timer.get_iterations())
# print(timer.get_iteration_time())
print("Tempo total: {}".format(timer.get_total_time()))
if best_fitness is not None:
best_fitness[-1] = global_best_fitness
# best_fitness[iteration] = np.mean(survival_values)
# best_fitness[iteration] = survival_values[-1]
if perf_counter is not None:
perf_counter[-1] = time.perf_counter() - start_perf_counter
if process_time is not None:
process_time[-1] = time.process_time() - start_process_time
# fitness_function(global_best_position, instance, timer, True)
return (global_best_position, global_best_fitness)
def prey_predator_algorithm(instance,
config,
fitness_function,
*,
best_fitness=None,
perf_counter=None,
process_time=None):
population_size = config.population_size
population = np.random.randint(2,
size=(population_size,
instance.num_materials),
dtype=bool)
timer = Timer(
) ########################################################################################################
start_perf_counter = time.perf_counter()
start_process_time = time.process_time()
for iteration in range(config.num_iterations):
timer.add_time(
) ########################################################################################################
# print('==========================' + str(iteration))
survival_values = fitness_function(population, instance, timer)
sorted_indices = np.argsort(survival_values)
population = population[sorted_indices]
survival_values = survival_values[sorted_indices]
# print('Survival values:\n{}\n'.format(survival_values))
if best_fitness is not None:
best_fitness[iteration] = survival_values[0]
# best_fitness[iteration] = np.mean(survival_values)
# best_fitness[iteration] = survival_values[-1]
if perf_counter is not None:
perf_counter[iteration] = time.perf_counter() - start_perf_counter
if process_time is not None:
process_time[iteration] = time.process_time() - start_process_time
new_population = np.copy(population)
timer.add_time(
"creation"
) ########################################################################################################
# Cria as mascaras para separar os diferentes tipos de individuo
best_prey_mask = np.zeros(population_size, dtype=bool)
best_prey_mask[0] = True
predator_mask = np.zeros(population_size, dtype=bool)
predator_mask[-1] = True
follow_mask = (np.random.rand(population_size) < config.follow_chance)
run_mask = ~follow_mask
follow_mask[best_prey_mask] = False # Ignora as melhores presas
follow_mask[predator_mask] = False # Ignora os predadores
run_mask[best_prey_mask] = False # Ignora as melhores presas
run_mask[predator_mask] = False # Ignora os predadores
# print('Best prey mask: {}'.format(best_prey_mask))
# print(' Predator mask: {}'.format(predator_mask))
# print(' Follow mask: {}'.format(follow_mask))
# print(' Run mask: {}'.format(run_mask))
timer.add_time(
) ########################################################################################################
follow_indices = np.where(follow_mask)[0]
follow_quant = len(follow_indices)
population_distance = [[
hamming_distance(population[j], population[i]) /
instance.num_materials for j in range(i)
] for i in follow_indices]
survival_ratio = [[
survival_values[j] / survival_values[i] for j in range(i)
] for i in follow_indices]
follow_chance = [[
(2 - config.follow_distance_parameter * population_distance[i][j] -
config.follow_survival_parameter * survival_ratio[i][j]) / 2
for j in range(follow_indices[i])
] for i in range(follow_quant)]
roulette_array = np.array(
[Roulette(follow_chance[i]) for i in range(follow_quant)])
timer.add_time(
"follow_chance"
) ########################################################################################################
# TODO(andre:2018-05-28): Garantir que max_steps nunca é maior do que o numero de materiais
num_steps = np.round(
config.max_steps * np.random.rand(follow_quant) /
np.exp(config.steps_distance_parameter *
np.array([i[-1] for i in population_distance])))
new_population[follow_mask] = move_population_roulette(
new_population[follow_mask], num_steps, roulette_array, population)
timer.add_time(
"follow_roulette"
) ########################################################################################################
# num_steps = np.round(config.min_steps * np.random.rand(population_size))
# new_population = move_population_random(new_population, num_steps, follow_mask)
num_steps = np.round(config.min_steps * np.random.rand(follow_quant))
new_population[follow_mask] = move_population_random(
new_population[follow_mask], num_steps)
timer.add_time(
"follow_random"
) ########################################################################################################
# num_steps = np.round(config.max_steps * np.random.rand(population_size))
# new_population = move_population_random_complement(new_population, num_steps, population[-1], run_mask)
num_steps = np.round(config.max_steps *
np.random.rand(np.count_nonzero(run_mask)))
new_population[run_mask] = move_population_random_complement(
new_population[run_mask], num_steps, population[-1])
timer.add_time(
"run"
) ########################################################################################################
new_population[best_prey_mask] = move_population_local_search(
new_population[best_prey_mask], fitness_function, config.min_steps,
config.local_search_tries, instance, timer)
timer.add_time(
) ########################################################################################################
# num_steps = np.round(config.max_steps * np.random.rand(population_size))
# new_population = move_population_random(new_population, num_steps, predator_mask)
num_steps = np.round(config.max_steps *
np.random.rand(np.count_nonzero(predator_mask)))
new_population[predator_mask] = move_population_random(
new_population[predator_mask], num_steps)
timer.add_time(
"predator_random"
) ########################################################################################################
# num_steps = np.round(config.min_steps * np.random.rand(population_size))
# worst_prey = np.repeat(population[-2][np.newaxis, :], population_size, axis=0)
# new_population = move_population_direction(new_population, num_steps, worst_prey, predator_mask)
num_steps = np.round(config.min_steps *
np.random.rand(np.count_nonzero(predator_mask)))
worst_prey = np.repeat(population[-2][np.newaxis, :],
np.count_nonzero(predator_mask),
axis=0)
new_population[predator_mask] = move_population_direction(
new_population[predator_mask], num_steps, worst_prey)
timer.add_time(
"predator_follow"
) ########################################################################################################
# new_survival_values = fitness_function(new_population, instance, timer)
# print('Old population:\n{}\n'.format(population))
# print('New population:\n{}\n'.format(new_population))
# print('Comparison:\n{}\n'.format(population == new_population))
# print('Old survival values:\n{}\n'.format(survival_values))
# print('New survival values:\n{}\n'.format(new_survival_values))
population = new_population
print("Tempo: ")
print(timer.get_time())
print("Iterações: ")
print(timer.get_iterations())
# print(timer.get_iteration_time())
print("Tempo total: {}".format(timer.get_total_time()))
survival_values = fitness_function(population, instance, timer)
sorted_indices = np.argsort(survival_values)
population = population[sorted_indices]
survival_values = survival_values[sorted_indices]
if best_fitness is not None:
best_fitness[-1] = survival_values[0]
# best_fitness[-1] = np.mean(survival_values)
# best_fitness[-1] = survival_values[-1]
if perf_counter is not None:
perf_counter[-1] = time.perf_counter() - start_perf_counter
if process_time is not None:
process_time[-1] = time.process_time() - start_process_time
return (population, survival_values)
def prey_predator_algorithm_continuous(instance, config, fitness_function, out_info=None, verbose=False):
population_size = config.population_size
if config.max_steps > instance.num_materials:
config.max_steps = instance.num_materials
if config.min_steps > config.max_steps:
config.min_steps = config.max_steps
if config.evaluator == Evaluator.FIXED_EVALUATOR:
evaluate_function = evaluate_population_fixed
else:
evaluate_function = evaluate_population_random
def counter_fitness(individual, instance, student, timer, print_results=False, data=None):
nonlocal cost_counter
cost_counter += 1
return fitness_function(individual, instance, student, timer, print_results, data=data)
if out_info is not None:
out_info["best_fitness"] = []
out_info["partial_fitness"] = []
out_info["perf_counter"] = []
out_info["process_time"] = []
out_info["cost_value"] = []
results = []
for student in range(instance.num_learners):
if verbose:
print('[PPAC] Students progress: %d / %d (%d%%)' % (student + 1, instance.num_learners, (student + 1) * 100 / instance.num_learners))
cost_counter = 0
iteration_counter = 0
stagnation_counter = 0
if out_info is not None:
out_info['best_fitness'].append([])
out_info['partial_fitness'].append([])
out_info['perf_counter'].append([])
out_info['process_time'].append([])
out_info['cost_value'].append([])
timer = Timer()
population = np.random.rand(population_size, instance.num_materials) * (2 * config.max_position) - config.max_position
population_best_evaluation = evaluate_function(population[0])
population_best_fitness = counter_fitness(population_best_evaluation, instance, student, timer)
start_perf_counter = time.perf_counter()
start_process_time = time.process_time()
while ((not config.cost_budget or cost_counter < config.cost_budget) and
(not config.num_iterations or iteration_counter < config.num_iterations) and
(not config.max_stagnation or stagnation_counter < config.max_stagnation)):
timer.add_time()
population_evaluation = evaluate_function(population)
survival_values = np.apply_along_axis(counter_fitness, 1, population_evaluation, instance, student, timer)
sorted_indices = np.argsort(survival_values)
population = population[sorted_indices]
survival_values = survival_values[sorted_indices]
iteration_counter += 1
if survival_values[0] < population_best_fitness:
population_best_evaluation = population_evaluation[sorted_indices[0]]
population_best_fitness = survival_values[0]
stagnation_counter = 0
else:
stagnation_counter += 1
if out_info is not None:
out_info["best_fitness"][-1].append(population_best_fitness)
fitness_function(population_best_evaluation, instance, student, timer, data=out_info["partial_fitness"][-1])
out_info["perf_counter"][-1].append(time.perf_counter() - start_perf_counter)
out_info["process_time"][-1].append(time.process_time() - start_process_time)
out_info["cost_value"][-1].append(cost_counter)
new_population = np.copy(population)
timer.add_time("creation")
# Cria as mascaras para separar os diferentes tipos de individuo
best_prey_mask = np.zeros(population_size, dtype=bool)
best_prey_mask[0] = True
predator_mask = np.zeros(population_size, dtype=bool)
predator_mask[-1] = True
follow_mask = (np.random.rand(population_size) < config.follow_chance)
run_mask = ~follow_mask
follow_mask[best_prey_mask] = False # Ignora as melhores presas
follow_mask[predator_mask] = False # Ignora os predadores
run_mask[best_prey_mask] = False # Ignora as melhores presas
run_mask[predator_mask] = False # Ignora os predadores
timer.add_time()
follow_indices = np.where(follow_mask)[0]
follow_quant = len(follow_indices)
tau = 1
follow_direction = np.empty((follow_quant, instance.num_materials))
for index in range(follow_quant):
i = follow_indices[index]
population_distance = vector_size(population - population[i])
survival_ratio = survival_values[i] / survival_values
population_direction = np.exp(survival_ratio ** tau - population_distance * 0.05)[:, np.newaxis] * (population - population[i])
individual_direction = np.sum(population_direction, axis=0)
normalized_direction = individual_direction / vector_size(individual_direction)
follow_direction[index] = normalized_direction
# Gerar direção multidimensional
# https://stackoverflow.com/questions/6283080/random-unit-vector-in-multi-dimensional-space
timer.add_time("follow_calculate_direction")
# TODO(andre:2018-05-28): Garantir que max_steps nunca é maior do que o numero de materiais
omega = .5
predator_distance = survival_values[-1] - survival_values[follow_mask]
num_steps = config.max_steps * np.random.rand(follow_quant) / np.exp(config.steps_distance_parameter * predator_distance ** omega)
new_population[follow_mask] = move_population_direction(new_population[follow_mask], num_steps, follow_direction)
timer.add_time("follow_direction")
num_steps = np.round(config.min_steps * np.random.rand(follow_quant))
new_population[follow_mask] = move_population_random(new_population[follow_mask], num_steps)
timer.add_time("follow_random")
num_steps = np.round(config.max_steps * np.random.rand(np.count_nonzero(run_mask)))
new_population[run_mask] = move_population_random_complement(new_population[run_mask], num_steps, population[-1])
timer.add_time("run")
new_population[best_prey_mask] = move_population_local_search(new_population[best_prey_mask], counter_fitness, evaluate_function, config.min_steps, config.local_search_tries, instance, student, timer)
timer.add_time()
# TODO(andre:2018-12-17): Gerar uma direção dentro de um circulo unitário
# e multiplicar por max_steps para determinar a direção aleatória
num_steps = np.round(config.max_steps * np.random.rand(np.count_nonzero(predator_mask)))
new_population[predator_mask] = move_population_random(new_population[predator_mask], num_steps)
timer.add_time("predator_random")
num_steps = np.round(config.min_steps * np.random.rand(np.count_nonzero(predator_mask)))
worst_prey = np.repeat(population[-2][np.newaxis, :], np.count_nonzero(predator_mask), axis=0)
new_population[predator_mask] = move_population_direction(new_population[predator_mask], num_steps, worst_prey)
timer.add_time("predator_follow")
new_population = np.clip(new_population, -config.max_position, config.max_position)
population = new_population
population_evaluation = evaluate_function(population)
survival_values = np.apply_along_axis(counter_fitness, 1, population_evaluation, instance, student, timer)
sorted_indices = np.argsort(survival_values)
population_evaluation = population_evaluation[sorted_indices]
survival_values = survival_values[sorted_indices]
if out_info is not None:
out_info["best_fitness"][-1].append(population_best_fitness)
fitness_function(population_best_evaluation, instance, student, timer, data=out_info["partial_fitness"][-1])
out_info["perf_counter"][-1].append(time.perf_counter() - start_perf_counter)
out_info["process_time"][-1].append(time.process_time() - start_process_time)
out_info["cost_value"][-1].append(cost_counter)
results.append((population_best_evaluation, population_best_fitness))
return results
def differential_evolution(instance, config, fitness_function, out_info=None, verbose=False):
population_size = config.population_size
if config.evaluator == Evaluator.FIXED_EVALUATOR:
evaluate_function = evaluate_population_fixed
else:
evaluate_function = evaluate_population_random
def counter_fitness(individual, instance, student, timer, print_results=False, data=None):
nonlocal cost_counter
cost_counter += 1
return fitness_function(individual, instance, student, timer, print_results, data=data)
if out_info is not None:
out_info["best_fitness"] = []
out_info["partial_fitness"] = []
out_info["perf_counter"] = []
out_info["process_time"] = []
out_info["cost_value"] = []
results = []
for student in range(instance.num_learners):
if verbose:
print('[DE] Students progress: %d / %d (%d%%)' % (student + 1, instance.num_learners, (student + 1) * 100 / instance.num_learners))
cost_counter = 0
iteration_counter = 0
stagnation_counter = 0
if out_info is not None:
out_info['best_fitness'].append([])
out_info['partial_fitness'].append([])
out_info['perf_counter'].append([])
out_info['process_time'].append([])
out_info['cost_value'].append([])
timer = Timer()
# TODO(andre: 2019-04-25): Testar utilizar um valor limite para os valores
# dos individuos, similar ao PSO e ao PPA_C
population = np.random.rand(population_size, instance.num_materials) * (2 * config.max_velocity) - config.max_velocity
population_evaluation = evaluate_function(population)
survival_values = np.apply_along_axis(counter_fitness, 1, population_evaluation, instance, student, timer)
population_best_index = np.argmin(survival_values, axis=0)
population_best_evaluation = np.copy(population_evaluation[population_best_index])
population_best_fitness = survival_values[population_best_index]
start_perf_counter = time.perf_counter()
start_process_time = time.process_time()
while ((not config.cost_budget or cost_counter < config.cost_budget) and
(not config.num_iterations or iteration_counter < config.num_iterations) and
(not config.max_stagnation or stagnation_counter < config.max_stagnation)):
timer.add_time()
old_population_best_fitness = population_best_fitness
sorted_indices = np.argsort(survival_values)
population = population[sorted_indices]
survival_values = survival_values[sorted_indices]
if out_info is not None:
out_info["best_fitness"][-1].append(population_best_fitness)
fitness_function(population_best_evaluation, instance, student, timer, data=out_info["partial_fitness"][-1])
out_info["perf_counter"][-1].append(time.perf_counter() - start_perf_counter)
out_info["process_time"][-1].append(time.process_time() - start_process_time)
out_info["cost_value"][-1].append(cost_counter)
new_population = np.copy(population)
#--de
for p in range(population_size):
idxs = [idx for idx in range(population_size) if idx != p]
a, b, c = population[np.random.choice(idxs, 3, replace = False)]
# mutant = np.clip(a + config.mutation_chance * (b - c), 0, 1)
# mutant = np.copy(a + config.mutation_chance * (b - c))
mutant = np.clip(a + config.mutation_chance * (b - c), -config.max_velocity, config.max_velocity)
cross_points = np.random.rand(instance.num_materials) < config.crossover_rate
if not np.any(cross_points):
cross_points[np.random.randint(0, instance.num_materials)] = True
applicant = np.where(cross_points, mutant, population[p])
applicant_evaluation = evaluate_function(applicant)
applicant_fit = counter_fitness(applicant_evaluation, instance, student, timer)
if applicant_fit < survival_values[p]:
new_population[p] = applicant
survival_values[p] = applicant_fit
if applicant_fit < population_best_fitness:
population_best_evaluation = applicant_evaluation
population_best_fitness = applicant_fit
#--end de
iteration_counter += 1
if population_best_fitness < old_population_best_fitness:
stagnation_counter = 0
else:
stagnation_counter += 1
population = new_population
if out_info is not None:
out_info["best_fitness"][-1].append(population_best_fitness)
fitness_function(population_best_evaluation, instance, student, timer, data=out_info["partial_fitness"][-1])
out_info["perf_counter"][-1].append(time.perf_counter() - start_perf_counter)
out_info["process_time"][-1].append(time.process_time() - start_process_time)
out_info["cost_value"][-1].append(cost_counter)
results.append((population_best_evaluation, population_best_fitness))
return results
def genetic_algorithm(instance, config, fitness_function, out_info=None):
population_size = config.population_size
def counter_fitness(individual,
instance,
student,
timer,
print_results=False,
data=None):
nonlocal cost_counter
cost_counter += 1
return fitness_function(individual,
instance,
student,
timer,
print_results,
data=data)
if out_info is not None:
out_info['best_fitness'] = []
out_info['partial_fitness'] = []
out_info['perf_counter'] = []
out_info['process_time'] = []
out_info['cost_value'] = []
results = []
for student in range(instance.num_learners):
print(
'\n\n\n\n---------------------------------------------\n\n\n\nNew Student'
)
cost_counter = 0
iteration_counter = 0
stagnation_counter = 0
if out_info is not None:
out_info['best_fitness'].append([])
out_info['partial_fitness'].append([])
out_info['perf_counter'].append([])
out_info['process_time'].append([])
out_info['cost_value'].append([])
timer = Timer()
population = np.random.randint(2,
size=(population_size,
instance.num_materials),
dtype=bool)
population_best_individual = population[0]
population_best_fitness = counter_fitness(population[0], instance,
student, timer)
start_perf_counter = time.perf_counter()
start_process_time = time.process_time()
while ((not config.cost_budget or cost_counter < config.cost_budget)
and (not config.num_iterations
or iteration_counter < config.num_iterations)
and (not config.max_stagnation
or stagnation_counter < config.max_stagnation)):
timer.add_time()
survival_values = np.apply_along_axis(counter_fitness, 1,
population, instance,
student, timer)
print('----------------------')
sorted_indices = np.argsort(survival_values)
population = population[sorted_indices]
survival_values = survival_values[sorted_indices]
if survival_values[0] < population_best_fitness:
population_best_individual = population[0]
population_best_fitness = survival_values[0]
stagnation_counter = 0
else:
stagnation_counter += 1
print(population_best_fitness)
iteration_counter += 1
if out_info is not None:
out_info['best_fitness'][-1].append(population_best_fitness)
fitness_function(population_best_individual,
instance,
student,
timer,
data=out_info['partial_fitness'][-1])
out_info['perf_counter'][-1].append(time.perf_counter() -
start_perf_counter)
out_info['process_time'][-1].append(time.process_time() -
start_process_time)
out_info['cost_value'][-1].append(cost_counter)
new_population = copying_gene(population, config.copying_method,
config)
if config.use_local_search:
new_population = local_search_gene(new_population,
counter_fitness,
config.local_search_method,
config)
remaining_spots = population_size - len(new_population)
selection_spots = remaining_spots
if (config.crossover_method == Crossover.THREE_PARENT_CROSSOVER):
selection_spots = int(3 * math.ceil(remaining_spots / 3.)) * 3
else:
selection_spots = int(2 * math.ceil(remaining_spots / 2.))
parents = selection_gene(population, survival_values,
selection_spots, config.selection_method,
config)
children = crossover_gene(parents, config.crossover_method, config)
mutated = mutation_gene(children, config.mutation_method, config)
new_population = np.append(new_population,
mutated[:remaining_spots],
axis=0)
population = new_population
if out_info is not None:
out_info['best_fitness'][-1].append(population_best_fitness)
fitness_function(population_best_individual,
instance,
student,
timer,
data=out_info['partial_fitness'][-1])
out_info['perf_counter'][-1].append(time.perf_counter() -
start_perf_counter)
out_info['process_time'][-1].append(time.process_time() -
start_process_time)
out_info['cost_value'][-1].append(cost_counter)
results.append((population_best_individual, population_best_fitness))
return results
def nsga_ii(instance, config, fitness_function, out_info=None, verbose=False, **kwargs):
population_size = config.population_size
def counter_fitness(individual, instance, student, timer, print_results=False, data=None, **kwargs):
nonlocal cost_counter
cost_counter += 1
return fitness_function(individual, instance, student, timer, print_results, data=data, **kwargs)
if out_info is not None:
out_info['best_fitness'] = []
out_info['population_fitness'] = []
out_info['perf_counter'] = []
out_info['process_time'] = []
out_info['cost_value'] = []
results = []
for student in range(instance.num_learners):
if verbose:
print('[NSGA-II] Students progress: %d / %d (%d%%)' % (student + 1, instance.num_learners, (student + 1) * 100 / instance.num_learners))
cost_counter = 0
iteration_counter = 0
stagnation_counter = 0
if out_info is not None:
out_info['best_fitness'].append([])
out_info['population_fitness'].append([])
out_info['perf_counter'].append([])
out_info['process_time'].append([])
out_info['cost_value'].append([])
timer = Timer()
population = np.random.randint(2, size=(population_size, instance.num_materials), dtype=bool)
population_best_fitness = sum(counter_fitness(population[0], instance, student, timer, **kwargs))
survival_values = np.apply_along_axis(counter_fitness, 1, population, instance, student, timer, **kwargs)
sorted_fronts = sort_nondominated(survival_values, sign=-1)
distance = np.zeros((survival_values.shape[0],))
for front in sorted_fronts:
distance[front] = crowding_dist(survival_values[front])
start_perf_counter = time.perf_counter()
start_process_time = time.process_time()
while ((not config.cost_budget or cost_counter < config.cost_budget) and
(not config.num_iterations or iteration_counter < config.num_iterations) and
(not config.max_stagnation or stagnation_counter < config.max_stagnation)):
timer.add_time()
cur_best = np.min(np.sum(survival_values, axis=1))
if cur_best < population_best_fitness:
population_best_fitness = cur_best
stagnation_counter = 0
else:
stagnation_counter += 1
iteration_counter += 1
if out_info is not None:
out_info['best_fitness'][-1].append(population_best_fitness)
out_info['population_fitness'][-1].append(survival_values)
out_info['perf_counter'][-1].append(time.perf_counter() - start_perf_counter)
out_info['process_time'][-1].append(time.process_time() - start_process_time)
out_info['cost_value'][-1].append(cost_counter)
remaining_spots = population_size
selection_spots = remaining_spots
if (config.crossover_method == Crossover.THREE_PARENT_CROSSOVER):
selection_spots = int(3 * math.ceil(remaining_spots / 3.)) * 3
else:
selection_spots = int(2 * math.ceil(remaining_spots / 2.))
parents = selection_gene(population, survival_values, selection_spots, Selection.NSGA_II_SELECTION, config, crowding_dist=distance)
children = crossover_gene(parents, config.crossover_method, config)
mutated = mutation_gene(children, config.mutation_method, config)
mutated = mutated[:remaining_spots]
# Calculates the survival value of only the new individuals
new_survival_values = np.apply_along_axis(counter_fitness, 1, mutated, instance, student, timer, **kwargs)
new_population = np.append(population, mutated, axis=0)
new_survival_values = np.append(survival_values, new_survival_values, axis=0)
sorted_fronts = sort_nondominated(new_survival_values, sign=-1)
new_distance = np.zeros((new_survival_values.shape[0],))
sorted_indices = []
for front in sorted_fronts:
front_distance = crowding_dist(new_survival_values[front])
sorted_indices.extend([x for (_, x) in sorted(zip(front_distance, front), reverse=True)])
new_distance[front] = front_distance
new_population = new_population[sorted_indices]
new_survival_values = new_survival_values[sorted_indices]
new_distance = new_distance[sorted_indices]
population = new_population[:population_size]
survival_values = new_survival_values[:population_size]
distance = new_distance[:population_size]
if out_info is not None:
out_info['best_fitness'][-1].append(population_best_fitness)
out_info['population_fitness'][-1].append(survival_values)
out_info['perf_counter'][-1].append(time.perf_counter() - start_perf_counter)
out_info['process_time'][-1].append(time.process_time() - start_process_time)
out_info['cost_value'][-1].append(cost_counter)
results.append((population, survival_values))
return results