def EO(num_agents, max_iter, train_data, train_label, obj_function=compute_fitness, trans_func_shape='s', save_conv_graph=False):
# Equilibrium Optimizer
############################### Parameters ####################################
# #
# num_agents: number of particles #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# trans_function_shape: shape of the transfer function used #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
short_name = 'EO'
agent_name = 'Particle'
train_data, train_label = np.array(train_data), np.array(train_label)
num_features = train_data.shape[1]
trans_function = get_trans_function(trans_func_shape)
# initialize particles and Leader (the agent with the max fitness)
particles = initialize(num_agents, num_features)
fitness = np.zeros(num_agents)
accuracy = np.zeros(num_agents)
Leader_agent = np.zeros((1, num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
pool_size = 4
omega = 0.9
a2 = 1
a1 = 2
GP = 0.5
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
convergence_curve['feature_count'] = np.zeros(max_iter)
# format the data
data = Data()
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(
train_data, train_label, stratify=train_label, test_size=0.2)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# rank initial particles
particles, fitness = sort_agents(particles, obj_function, data)
# start timer
start_time = time.time()
# pool initialization
eq_pool = np.zeros((pool_size+1, num_features))
eq_fitness = np.zeros(pool_size)
eq_fitness[:] = float("-inf")
for iter_no in range(max_iter):
print('\n================================================================================')
print(' Iteration - {}'.format(iter_no+1))
print('================================================================================\n')
# replacements in the pool
for i in range(num_agents):
for j in range(pool_size):
if fitness[i] <= eq_fitness[j]:
eq_fitness[j] = fitness[i].copy()
eq_pool[j, :] = particles[i, :].copy()
break
best_particle = eq_pool[0,:]
Cave = avg_concentration(eq_pool, pool_size, num_features)
eq_pool[pool_size] = Cave.copy()
t = (1 - (iter_no/max_iter)) ** (a2*iter_no/max_iter)
for i in range(num_agents):
# randomly choose one candidate from the equillibrium pool
inx = np.random.randint(0,pool_size)
Ceq = np.array(eq_pool[inx])
lambda_vec = np.zeros(np.shape(Ceq))
r_vec = np.zeros(np.shape(Ceq))
for j in range(num_features):
lambda_vec[j] = np.random.random()
r_vec[j] = np.random.random()
F_vec = np.zeros(np.shape(Ceq))
for j in range(num_features):
x = -1*lambda_vec[j]*t
x = np.exp(x) - 1
x = a1 * sign_func(r_vec[j] - 0.5) * x
r1, r2 = np.random.random(2)
if r2 < GP:
GCP = 0
else:
GCP = 0.5 * r1
G0 = np.zeros(np.shape(Ceq))
G = np.zeros(np.shape(Ceq))
for j in range(num_features):
G0[j] = GCP * (Ceq[j] - lambda_vec[j]*particles[i][j])
G[j] = G0[j]*F_vec[j]
# use transfer function to map continuous->binary
for j in range(num_features):
temp = Ceq[j] + (particles[i][j] - Ceq[j])*F_vec[j] + G[j]*(1 - F_vec[j])/lambda_vec[j]
temp = trans_function(temp)
if temp>np.random.random():
particles[i][j] = 1 - particles[i][j]
else:
particles[i][j] = particles[i][j]
# update final information
particles, fitness = sort_agents(particles, obj_function, data)
display(particles, fitness, agent_name)
# update Leader (best agent)
if fitness[0] > Leader_fitness:
Leader_agent = particles[0].copy()
Leader_fitness = fitness[0].copy()
convergence_curve['fitness'][iter_no] = Leader_fitness
convergence_curve['feature_count'][iter_no] = int(np.sum(Leader_agent))
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy, data)
particles, accuracy = sort_agents(particles, compute_accuracy, data)
print('\n================================================================================')
print(' Final Result ')
print('================================================================================\n')
print('Leader ' + agent_name + ' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name + ' Classification Accuracy : {}'.format(Leader_accuracy))
print('\n================================================================================\n')
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence curves
iters = np.arange(max_iter)+1
fig, axes = plt.subplots(2, 1)
fig.tight_layout(pad=5)
fig.suptitle('Convergence Curves')
axes[0].set_title('Convergence of Fitness over Iterations')
axes[0].set_xlabel('Iteration')
axes[0].set_ylabel('Fitness')
axes[0].plot(iters, convergence_curve['fitness'])
axes[1].set_title('Convergence of Feature Count over Iterations')
axes[1].set_xlabel('Iteration')
axes[1].set_ylabel('Number of Selected Features')
axes[1].plot(iters, convergence_curve['feature_count'])
if(save_conv_graph):
plt.savefig('convergence_graph_'+ short_name + '.jpg')
plt.show()
# update attributes of solution
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_particles = particles
solution.final_fitness = fitness
solution.final_accuracy = accuracy
solution.execution_time = exec_time
return solution
def SCA(num_agents,
max_iter,
train_data,
train_label,
obj_function=compute_fitness,
trans_func_shape='s',
save_conv_graph=False):
# Sine Cosine Algorithm
############################### Parameters ####################################
# #
# num_agents: number of agents #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# trans_function_shape: shape of the transfer function used #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
short_name = 'SCA'
agent_name = 'Agent'
train_data, train_label = np.array(train_data), np.array(train_label)
num_features = train_data.shape[1]
trans_function = get_trans_function(trans_func_shape)
# setting up the objectives
weight_acc = None
if (obj_function == compute_fitness):
weight_acc = float(
input('Weight for the classification accuracy [0-1]: '))
obj = (obj_function, weight_acc)
compute_accuracy = (
compute_fitness, 1
) # compute_accuracy is just compute_fitness with accuracy weight as 1
# initialize agents and Leader (the agent with the max fitness)
population = initialize(num_agents, num_features)
fitness = np.zeros(num_agents)
accuracy = np.zeros(num_agents)
Leader_agent = np.zeros((1, num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
convergence_curve['feature_count'] = np.zeros(max_iter)
# initialize data class
data = Data()
val_size = float(
input('Enter the percentage of data wanted for valdiation [0, 100]: ')
) / 100
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(
train_data, train_label, stratify=train_label, test_size=val_size)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# rank initial population
population, fitness = sort_agents(population, obj, data)
Leader_agent = population[0].copy()
Leader_fitness = fitness[0].copy()
# start timer
start_time = time.time()
# Eq. (3.4)
a = 3
for iter_no in range(max_iter):
print(
'\n================================================================================'
)
print(' Iteration - {}'.format(iter_no + 1))
print(
'================================================================================\n'
)
# Eq. (3.4)
r1 = a - iter_no * (
(a) / max_iter) # r1 decreases linearly from a to 0
# update the Position of search agents
for i in range(num_agents):
for j in range(num_features):
# update r2, r3, and r4 for Eq. (3.3)
r2 = (2 * np.pi) * np.random.random()
r3 = 2 * np.random.random()
r4 = np.random.random()
# Eq. (3.3)
if r4 < 0.5:
# Eq. (3.1)
population[i, j] = population[i, j] + \
(r1*np.sin(r2)*abs(r3*Leader_agent[j]-population[i, j]))
else:
# Eq. (3.2)
population[i, j] = population[i, j] + \
(r1*np.cos(r2)*abs(r3*Leader_agent[j]-population[i, j]))
temp = population[i, j].copy()
temp = trans_function(temp)
if temp > np.random.random():
population[i, j] = 1
else:
population[i, j] = 0
# update final information
population, fitness = sort_agents(population, obj, data)
display(population, fitness)
if fitness[0] > Leader_fitness:
Leader_agent = population[0].copy()
Leader_fitness = fitness[0].copy()
convergence_curve['fitness'][iter_no] = Leader_fitness
convergence_curve['feature_count'][iter_no] = int(np.sum(Leader_agent))
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy,
data)
population, accuracy = sort_agents(population, compute_accuracy, data)
print(
'\n================================================================================'
)
print(
' Final Result '
)
print(
'================================================================================\n'
)
print('Leader ' + agent_name +
' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name +
' Classification Accuracy : {}'.format(Leader_accuracy))
print(
'\n================================================================================\n'
)
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence curves
iters = np.arange(max_iter) + 1
fig, axes = plt.subplots(2, 1)
fig.tight_layout(pad=5)
fig.suptitle('Convergence Curves')
axes[0].set_title('Convergence of Fitness over Iterations')
axes[0].set_xlabel('Iteration')
axes[0].set_ylabel('Fitness')
axes[0].plot(iters, convergence_curve['fitness'])
axes[1].set_title('Convergence of Feature Count over Iterations')
axes[1].set_xlabel('Iteration')
axes[1].set_ylabel('Number of Selected Features')
axes[1].plot(iters, convergence_curve['feature_count'])
if (save_conv_graph):
plt.savefig('convergence_graph_' + short_name + '.jpg')
plt.show()
# update attributes of solution
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_particles = population
solution.final_fitness = fitness
solution.final_accuracy = accuracy
solution.execution_time = exec_time
return solution
def RDA(num_agents, max_iter, train_data, train_label, obj_function=compute_fitness, trans_function_shape='s', save_conv_graph=False):
# Red Deer Algorithm
############################### Parameters ####################################
# #
# num_agents: number of red deers #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# trans_function_shape: shape of the transfer function used #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
# Number of agents must be at least 8
if num_agents < 8:
print("The parameter num_agents must be at least 8", file=sys.stderr)
sys.exit()
short_name = 'RDA'
agent_name = 'RedDeer'
train_data, train_label = np.array(train_data), np.array(train_label)
num_features = train_data.shape[1]
trans_function = get_trans_function(trans_function_shape)
# initialize red deers and Leader (the agent with the max fitness)
deer = initialize(num_agents, num_features)
fitness = np.zeros(num_agents)
accuracy = np.zeros(num_agents)
Leader_agent = np.zeros((1, num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
convergence_curve['feature_count'] = np.zeros(max_iter)
# format the data
data = Data()
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(train_data, train_label, stratify=train_label, test_size=0.2)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# initializing parameters
UB = 5 # Upper bound
LB = -5 # Lower bound
gamma = 0.5 # Fraction of total number of males who are chosen as commanders
alpha = 0.2 # Fraction of total number of hinds in a harem who mate with the commander of their harem
beta = 0.1 # Fraction of total number of hinds in a harem who mate with the commander of a different harem
# start timer
start_time = time.time()
# main loop
for iter_no in range(max_iter):
print('\n================================================================================')
print(' Iteration - {}'.format(iter_no+1))
print('================================================================================\n')
deer, fitness = sort_agents(deer, obj_function, data)
num_males = int(0.25 * num_agents)
num_hinds = num_agents - num_males
males = deer[:num_males,:]
hinds = deer[num_males:,:]
# roaring of male deer
for i in range(num_males):
r1 = np.random.random() # r1 is a random number in [0, 1]
r2 = np.random.random() # r2 is a random number in [0, 1]
r3 = np.random.random() # r3 is a random number in [0, 1]
new_male = males[i].copy()
if r3 >= 0.5: # Eq. (3)
new_male += r1 * (((UB - LB) * r2) + LB)
else:
new_male -= r1 * (((UB - LB) * r2) + LB)
# apply transformation function on the new male
for j in range(num_features):
trans_value = trans_function(new_male[j])
if (np.random.random() < trans_value):
new_male[j] = 1
else:
new_male[j] = 0
if obj_function(new_male, data.train_X, data.val_X, data.train_Y, data.val_Y) < obj_function(males[i], data.train_X, data.val_X, data.train_Y, data.val_Y):
males[i] = new_male
# selection of male commanders and stags
num_coms = int(num_males * gamma) # Eq. (4)
num_stags = num_males - num_coms # Eq. (5)
coms = males[:num_coms,:]
stags = males[num_coms:,:]
# fight between male commanders and stags
for i in range(num_coms):
chosen_com = coms[i].copy()
chosen_stag = random.choice(stags)
r1 = np.random.random()
r2 = np.random.random()
new_male_1 = (chosen_com + chosen_stag) / 2 + r1 * (((UB - LB) * r2) + LB) # Eq. (6)
new_male_2 = (chosen_com + chosen_stag) / 2 - r1 * (((UB - LB) * r2) + LB) # Eq. (7)
# apply transformation function on new_male_1
for j in range(num_features):
trans_value = trans_function(new_male_1[j])
if (np.random.random() < trans_value):
new_male_1[j] = 1
else:
new_male_1[j] = 0
# apply transformation function on new_male_2
for j in range(num_features):
trans_value = trans_function(new_male_2[j])
if (np.random.random() < trans_value):
new_male_2[j] = 1
else:
new_male_2[j] = 0
fitness = np.zeros(4)
fitness[0] = obj_function(chosen_com, data.train_X, data.val_X, data.train_Y, data.val_Y)
fitness[1] = obj_function(chosen_stag, data.train_X, data.val_X, data.train_Y, data.val_Y)
fitness[2] = obj_function(new_male_1, data.train_X, data.val_X, data.train_Y, data.val_Y)
fitness[3] = obj_function(new_male_2, data.train_X, data.val_X, data.train_Y, data.val_Y)
bestfit = np.max(fitness)
if fitness[0] < fitness[1] and fitness[1] == bestfit:
coms[i] = chosen_stag.copy()
elif fitness[0] < fitness[2] and fitness[2] == bestfit:
coms[i] = new_male_1.copy()
elif fitness[0] < fitness[3] and fitness[3] == bestfit:
coms[i] = new_male_2.copy()
# formation of harems
coms, fitness = sort_agents(coms, obj_function, data)
norm = np.linalg.norm(fitness)
normal_fit = fitness / norm
total = np.sum(normal_fit)
power = normal_fit / total # Eq. (9)
num_harems = [int(x * num_hinds) for x in power] # Eq.(10)
max_harem_size = np.max(num_harems)
harem = np.empty(shape=(num_coms, max_harem_size, num_features))
random.shuffle(hinds)
itr = 0
for i in range(num_coms):
harem_size = num_harems[i]
for j in range(harem_size):
harem[i][j] = hinds[itr]
itr += 1
# mating of commander with hinds in his harem
num_harem_mate = [int(x * alpha) for x in num_harems] # Eq. (11)
population_pool = list(deer)
for i in range(num_coms):
random.shuffle(harem[i])
for j in range(num_harem_mate[i]):
r = np.random.random() # r is a random number in [0, 1]
offspring = (coms[i] + harem[i][j]) / 2 + (UB - LB) * r # Eq. (12)
# apply transformation function on offspring
for j in range(num_features):
trans_value = trans_function(offspring[j])
if (np.random.random() < trans_value):
offspring[j] = 1
else:
offspring[j] = 0
population_pool.append(list(offspring))
# if number of commanders is greater than 1, inter-harem mating takes place
if num_coms > 1:
# mating of commander with hinds in another harem
k = i
while k == i:
k = random.choice(range(num_coms))
num_mate = int(num_harems[k] * beta) # Eq. (13)
np.random.shuffle(harem[k])
for j in range(num_mate):
r = np.random.random() # r is a random number in [0, 1]
offspring = (coms[i] + harem[k][j]) / 2 + (UB - LB) * r
# apply transformation function on offspring
for j in range(num_features):
trans_value = trans_function(offspring[j])
if (np.random.random() < trans_value):
offspring[j] = 1
else:
offspring[j] = 0
population_pool.append(list(offspring))
# mating of stag with nearest hind
for stag in stags:
dist = np.zeros(num_hinds)
for i in range(num_hinds):
dist[i] = math.sqrt(np.sum((stag-hinds[i])*(stag-hinds[i])))
min_dist = np.min(dist)
for i in range(num_hinds):
distance = math.sqrt(np.sum((stag-hinds[i])*(stag-hinds[i]))) # Eq. (14)
if(distance == min_dist):
r = np.random.random() # r is a random number in [0, 1]
offspring = (stag + hinds[i])/2 + (UB - LB) * r
# apply transformation function on offspring
for j in range(num_features):
trans_value = trans_function(offspring[j])
if (np.random.random() < trans_value):
offspring[j] = 1
else:
offspring[j] = 0
population_pool.append(list(offspring))
break
# selection of the next generation
population_pool = np.array(population_pool)
population_pool, fitness = sort_agents(population_pool, obj_function, data)
maximum = sum([f for f in fitness])
selection_probs = [f/maximum for f in fitness]
indices = np.random.choice(len(population_pool), size=num_agents, replace=True, p=selection_probs)
deer = population_pool[indices]
# update final information
deer, fitness = sort_agents(deer, obj_function, data)
display(deer, fitness, agent_name)
if fitness[0] > Leader_fitness:
Leader_agent = deer[0].copy()
Leader_fitness = fitness[0].copy()
convergence_curve['fitness'][iter_no] = Leader_fitness
convergence_curve['feature_count'][iter_no] = int(np.sum(Leader_agent))
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy, data)
deer, accuracy = sort_agents(deer, compute_accuracy, data)
print('\n================================================================================')
print(' Final Result ')
print('================================================================================\n')
print('Leader ' + agent_name + ' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name + ' Classification Accuracy : {}'.format(Leader_accuracy))
print('\n================================================================================\n')
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence curves
iters = np.arange(max_iter)+1
fig, axes = plt.subplots(2, 1)
fig.tight_layout(pad = 5)
fig.suptitle('Convergence Curves')
axes[0].set_title('Convergence of Fitness over Iterations')
axes[0].set_xlabel('Iteration')
axes[0].set_ylabel('Fitness')
axes[0].plot(iters, convergence_curve['fitness'])
axes[1].set_title('Convergence of Feature Count over Iterations')
axes[1].set_xlabel('Iteration')
axes[1].set_ylabel('Number of Selected Features')
axes[1].plot(iters, convergence_curve['feature_count'])
if(save_conv_graph):
plt.savefig('convergence_graph_'+ short_name + '.jpg')
plt.show()
# update attributes of solution
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_population = deer
solution.final_fitness = fitness
solution.final_accuracy = accuracy
solution.execution_time = exec_time
return solution
def GSA(num_agents,
max_iter,
train_data,
train_label,
obj_function=compute_fitness,
trans_function_shape='s',
save_conv_graph=False):
# Gravitational Search Algorithm
############################### Parameters ####################################
# #
# num_agents: number of particles #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# trans_function_shape: shape of the transfer function used #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
agent_name = 'Particle'
short_name = 'GSA'
train_data, train_label = np.array(train_data), np.array(train_label)
num_features = train_data.shape[1]
trans_function = get_trans_function(trans_function_shape)
# setting up the objectives
weight_acc = None
if (obj_function == compute_fitness):
weight_acc = float(
input('Weight for the classification accuracy [0-1]: '))
obj = (obj_function, weight_acc)
compute_accuracy = (
compute_fitness, 1
) # compute_accuracy is just compute_fitness with accuracy weight as 1
# initialize positionss of particles and Leader (the agent with the max fitness)
positions = initialize(num_agents, num_features)
fitness = np.zeros(num_agents)
accuracy = np.zeros(num_agents)
Leader_agent = np.zeros((1, num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
# initialize data class
data = Data()
val_size = float(
input('Enter the percentage of data wanted for valdiation [0, 100]: ')
) / 100
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(
train_data, train_label, stratify=train_label, test_size=val_size)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# initializing parameters
eps = 0.00001
G_ini = 6
F = np.zeros((num_agents, num_agents, num_features))
R = np.zeros((num_agents, num_agents))
force = np.zeros((num_agents, num_features))
acc = np.zeros((num_agents, num_features))
velocity = np.zeros((num_agents, num_features))
kBest = range(5)
# rank initial population
positions, fitness = sort_agents(positions, obj, data)
# start timer
start_time = time.time()
# main loop
for iter_no in range(max_iter):
print(
'\n================================================================================'
)
print(' Iteration - {}'.format(iter_no + 1))
print(
'================================================================================\n'
)
# updating value of G
G = G_ini - iter_no * (G_ini / max_iter) # Eq. (13)
# finding mass of each particle
best_fitness = fitness[0]
worst_fitness = fitness[-1]
m = (fitness - worst_fitness) / (best_fitness - worst_fitness + eps
) # Eq. (15)
sum_fitness = np.sum(m)
mass = m / sum_fitness # Eq. (16)
# finding force acting between each pair of particles
for i in range(num_agents):
for j in range(num_agents):
for k in range(num_features):
R[i][j] += abs(positions[i][k] -
positions[j][k]) # Eq. (8)
F[i][j] = G * (mass[i] * mass[j]) / (R[i][j] + eps) * (
positions[j] - positions[i]) # Eq. (7)
# finding net force acting on each particle
for i in range(num_agents):
for j in kBest:
if i != j:
force[i] += np.random.random() * F[i][j] # Eq. (9)
# finding acceleration of each particle
for i in range(num_agents):
acc[i] = force[i] / (mass[i] + eps) # Eq. (10)
# updating velocity of each particle
velocity = np.random.random() * velocity + acc # Eq. (11)
# apply transformation function on the velocity
for i in range(num_agents):
for j in range(num_features):
trans_value = trans_function(velocity[i][j])
if (np.random.random() < trans_value):
positions[i][j] = 1
else:
positions[i][j] = 0
if np.sum(positions[i]) == 0:
x = np.random.randint(0, num_features - 1)
positions[i][x] = 1
# update final information
positions, fitness = sort_agents(positions, obj, data)
display(positions, fitness, agent_name)
if fitness[0] > Leader_fitness:
Leader_agent = positions[0].copy()
Leader_fitness = fitness[0].copy()
convergence_curve['fitness'][iter_no] = np.mean(fitness)
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy,
data)
positions, accuracy = sort_agents(positions, compute_accuracy, data)
print(
'\n================================================================================'
)
print(
' Final Result '
)
print(
'================================================================================\n'
)
print('Leader ' + agent_name +
' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name +
' Classification Accuracy : {}'.format(Leader_accuracy))
print(
'\n================================================================================\n'
)
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence graph
fig, axes = Conv_plot(convergence_curve)
if (save_conv_graph):
plt.savefig('convergence_graph_' + short_name + '.jpg')
plt.show()
# update attributes of solution
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_population = positions
solution.final_fitness = fitness
solution.final_accuracy = accuracy
solution.execution_time = exec_time
return solution
def BBA(num_agents,
max_iter,
train_data,
train_label,
obj_function=compute_fitness,
trans_function_shape='s',
constantLoudness=True,
save_conv_graph=False):
# Binary Bat Algorithm (BBA)
############################### Parameters ####################################
# #
# num_agents: number of bats #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# trans_function_shape: shape of the transfer function used #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
agent_name = "Bat"
short_name = "BBA"
train_data, train_label = np.array(train_data), np.array(train_label)
num_features = train_data.shape[1]
trans_function = get_trans_function(trans_function_shape)
# setting up the objectives
weight_acc = None
if (obj_function == compute_fitness):
weight_acc = float(
input('Weight for the classification accuracy [0-1]: '))
obj = (obj_function, weight_acc)
compute_accuracy = (
compute_fitness, 1
) # compute_accuracy is just compute_fitness with accuracy weight as 1
# initialize bats and Leader (the agent with the max fitness)
bats = initialize(num_agents, num_features)
velocity = np.zeros([num_agents, num_features])
fitness = np.zeros(num_agents)
Leader_agent = np.zeros((1, num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
accuracy = np.zeros(num_agents)
# initialize some important parameters
minFrequency = 0 # min freq, const if constantLoudness == True
maxFrequency = 2 # max freq, const if constantLoudness == True
A = 1.00 # loudness
r = 0.15 # pulse emission rate
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
# initialize data class
data = Data()
val_size = float(
input('Enter the percentage of data wanted for valdiation [0, 100]: ')
) / 100
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(
train_data, train_label, stratify=train_label, test_size=val_size)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# start timer
start_time = time.time()
bats, fitness = sort_agents(bats, obj, data)
Leader_agent = bats[0, :]
Leader_fitness = fitness[0]
alpha = 0.95
gamma = 0.5
A_t = A
r_t = r
for iter_no in range(max_iter):
print(
'\n================================================================================'
)
print(' Iteration - {}'.format(iter_no + 1))
print(
'================================================================================\n'
)
if constantLoudness == False:
A_t *= alpha
r_t = r * (1 - np.exp(-1 * gamma * iter_no))
for agentNumber in range(num_agents):
fi = minFrequency + (maxFrequency - minFrequency) * np.random.rand(
) # frequency for i-th agent or bat
# update velocity equation number 1 in paper
velocity[agentNumber, :] = velocity[
agentNumber, :] + (bats[agentNumber, :] - Leader_agent) * fi
# updating the bats for bat number = agentNumber
newPos = np.zeros([1, num_features])
for featureNumber in range(num_features):
transferValue = trans_function(velocity[agentNumber,
featureNumber])
# change complement bats value at dimension number = featureNumber
if np.random.rand() < transferValue:
newPos[0, featureNumber] = 1 - bats[agentNumber,
featureNumber]
else:
newPos[0, featureNumber] = bats[agentNumber, featureNumber]
# considering the current pulse rate
if np.random.rand() > r_t:
newPos[0, featureNumber] = Leader_agent[featureNumber]
## calculate fitness for new bats
newFit = obj_function(newPos, data.train_X, data.val_X,
data.train_Y, data.val_Y, weight_acc)
## update better solution for indivisual bat
if fitness[agentNumber] <= newFit and np.random.rand() <= A_t:
fitness[agentNumber] = newFit
bats[agentNumber, :] = newPos[0, :]
bats, fitness = sort_agents(bats, obj, data)
## update (global) best solution for all bats
if fitness[0] > Leader_fitness:
Leader_fitness = fitness[0]
Leader_agent = bats[0, :]
convergence_curve['fitness'][iter_no] = np.mean(fitness)
# display current agents
display(bats, fitness, agent_name)
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy,
data)
bats, accuracy = sort_agents(bats, compute_accuracy, data)
print(
'\n================================================================================'
)
print(
' Final Result '
)
print(
'================================================================================\n'
)
print('Leader ' + agent_name +
' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name +
' Classification Accuracy : {}'.format(Leader_accuracy))
print(
'\n================================================================================\n'
)
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence graph
fig, axes = Conv_plot(convergence_curve)
if (save_conv_graph):
plt.savefig('convergence_graph_' + short_name + '.jpg')
plt.show()
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_population = bats
solution.final_fitness = fitness
solution.final_accuracy = accuracy
solution.execution_time = exec_time
return solution
def HS(num_agents, max_iter, train_data, train_label, obj_function = compute_fitness, save_conv_graph = False):
# Harmony Search Algorithm
############################### Parameters ####################################
# #
# num_agents: number of harmonies #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# trans_function_shape: shape of the transfer function used #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
# <STEPS OF HARMOMY SEARCH ALGORITH>
# Step 1. Initialize a Harmony Memory (HM).
# Step 2. Improvise a new harmony from HM.
# Step 3. If the new harmony is better than minimum harmony in HM, include the new harmony in HM, and exclude the minimum harmony from HM.
# Step 4. If stopping criteria are not satisfied, go to Step 2.
short_name = 'HS'
agent_name = 'Harmony'
train_data, train_label = np.array(train_data), np.array(train_label)
num_features = train_data.shape[1]
# intialize the harmonies and Leader (the agent with the max fitness)
harmonyMemory = initialize(num_agents, num_features)
fitness = np.zeros(num_agents)
accuracy = np.zeros(num_agents)
Leader_agent = np.zeros((1, num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
HMCR = 0.90 # Harmony Memory Consideration Rate
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
convergence_curve['feature_count'] = np.zeros(max_iter)
# format the data
data = Data()
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(train_data, train_label, stratify=train_label, test_size=0.2)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# start timer
start_time = time.time()
# calculate initial fitess and sort the harmony memory and rank them
harmonyMemory, fitness = sort_agents(harmonyMemory, obj_function, data)
# create new harmonies in each iteration
for iterCount in range(max_iter):
print('\n================================================================================')
print(' Iteration - {}'.format(iterCount + 1))
print('================================================================================\n')
HMCR_randValue = np.random.rand()
newHarmony = np.zeros([1, num_features])
# print(HMCR)
# print(HMCR_randValue)
if HMCR_randValue <= HMCR:
for featureNum in range(num_features):
selectedAgent = random.randint(0, num_agents - 1)
newHarmony[0, featureNum] = harmonyMemory[selectedAgent, featureNum]
else:
for featureNum in range(num_features):
newHarmony[0, featureNum] = random.randint(0, 1)
fitnessHarmony = obj_function(newHarmony, data.train_X, data.val_X, data.train_Y, data.val_Y)
if fitness[num_agents-1] < fitnessHarmony:
harmonyMemory[num_agents-1, :] = newHarmony
fitness[num_agents-1] = fitnessHarmony
# sort harmony memory
harmonyMemory, fitness = sort_agents(harmonyMemory, obj_function, data)
if fitness[0] > Leader_fitness:
Leader_agent = harmonyMemory[0].copy()
Leader_fitness = fitness[0].copy()
# update
convergence_curve['fitness'][iterCount] = Leader_fitness
convergence_curve['feature_count'][iterCount] = int(np.sum(Leader_agent))
display(harmonyMemory, fitness, agent_name)
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy, data)
harmonyMemory, accuracy = sort_agents(harmonyMemory, compute_accuracy, data)
print('\n================================================================================')
print(' Final Result ')
print('================================================================================\n')
print('Leader ' + agent_name + ' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name + ' Classification Accuracy : {}'.format(Leader_accuracy))
print('\n================================================================================\n')
# leader agent and leader fitneess
Leader_fitness = fitness[0]
Leader_agent = harmonyMemory[0].copy()
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence curves
iters = np.arange(max_iter)+1
fig, axes = plt.subplots(2, 1)
fig.tight_layout(pad=5)
fig.suptitle('Convergence Curves')
axes[0].set_title('Convergence of Fitness over Iterations')
axes[0].set_xlabel('Iteration')
axes[0].set_ylabel('Fitness')
axes[0].plot(iters, convergence_curve['fitness'])
axes[1].set_title('Convergence of Feature Count over Iterations')
axes[1].set_xlabel('Iteration')
axes[1].set_ylabel('Number of Selected Features')
axes[1].plot(iters, convergence_curve['feature_count'])
if(save_conv_graph):
plt.savefig('convergence_graph_'+ short_name + '.jpg')
plt.show()
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_population = harmonyMemory
solution.final_fitness = fitness
solution.final_accuracy = accuracy
solution.execution_time = exec_time
return solution
def MA(num_agents, max_iter, train_data, train_label, obj_function=compute_fitness, trans_function_shape='s', prob_mut=0.2, save_conv_graph=False):
# Mayfly Algorithm
############################### Parameters ####################################
# #
# num_agents: number of mayflies #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# prob_mut: probability of mutation #
# trans_function_shape: shape of the transfer function used #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
short_name = 'MA'
agent_name = 'Mayfly'
train_data, train_label = np.array(train_data), np.array(train_label)
num_features = train_data.shape[1]
trans_function = get_trans_function(trans_function_shape)
# control parameters
a1 = 1
a2 = 1.5
d = 0.1
fl = 0.1
g = 0.8
beta = 2
delta = 0.9
# initialize position and velocities of male and female mayflies' and Leader (the agent with the max fitness)
male_pos = initialize(num_agents, num_features)
female_pos = initialize(num_agents, num_features)
male_vel = np.random.uniform(low = -1, high = -1, size = (num_agents, num_features))
female_vel = np.random.uniform(low = -1, high = -1, size = (num_agents, num_features))
male_fitness = np.zeros((num_agents))
male_accuracy = np.zeros(num_agents)
female_fitness = np.zeros((num_agents))
Leader_agent = np.zeros((num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
male_personal_best = np.zeros((num_agents, num_features))
male_offspring = np.zeros((num_agents, num_features))
female_offspring = np.zeros((num_agents, num_features))
vmax_male = np.zeros((num_features))
vmax_female = np.zeros((num_features))
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
convergence_curve['feature_count'] = np.zeros(max_iter)
data = Data()
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(train_data, train_label, stratify=train_label, test_size=0.2)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# rank initial population
male_pos, male_fitness = sort_agents(male_pos, obj_function, data)
female_pos, female_fitness = sort_agents(female_pos, obj_function, data)
# start timer
start_time = time.time()
# main loop
for iter_no in range(max_iter):
print('\n================================================================================')
print(' Iteration - {}'.format(iter_no+1))
print('================================================================================\n')
#updating velocity limits
vmax_male, vmax_female = update_max_velocity(male_pos, female_pos)
for agent in range(num_agents):
#updating Leader fitness and personal best fitnesses
if male_fitness[agent] > Leader_fitness:
Leader_fitness = male_fitness[agent]
Leader_agent = male_pos[agent]
if male_fitness[agent] > obj_function(male_personal_best[agent], data.train_X, data.val_X, data.train_Y, data.val_Y):
male_personal_best[agent] = male_pos[agent]
#update velocities of male and female mayflies
male_vel[agent], female_vel[agent] = update_velocity(male_pos[agent], female_pos[agent], male_vel[agent], female_vel[agent], Leader_agent, male_personal_best[agent], a1, a2, d, fl, g, beta, agent, data, obj_function)
#check boundary condition of velocities of male and female mayflies
male_vel[agent], female_vel[agent] = check_velocity_limits(male_vel[agent], female_vel[agent], vmax_male, vmax_female)
#applying transfer functions to update positions of male and female mayflies
#the updation is done based on their respective velocity values
for j in range(num_features):
trans_value = trans_function(male_vel[agent][j])
if trans_value > np.random.normal(0,1):
male_pos[agent][j]=1
else:
male_pos[agent][j]=0
trans_value = trans_function(female_vel[agent][j])
if trans_value > np.random.random():
female_pos[agent][j]=1
else:
female_pos[agent][j]=0
#sorting
male_pos, male_fitness = sort_agents(male_pos, obj_function, data)
female_pos, female_fitness = sort_agents(female_pos, obj_function, data)
for agent in range(num_agents):
#generation of offsprings by crossover and mutation between male and female parent mayflies
male_offspring[agent], female_offspring[agent] = cross_mut(male_pos[agent], female_pos[agent])
#comparing parents and offsprings and replacing parents wherever necessary
male_pos = compare_and_replace(male_pos, male_offspring, male_fitness, data, obj_function)
female_pos = compare_and_replace(female_pos, female_offspring, female_fitness, data, obj_function)
#updating fitness values
male_pos, male_fitness = sort_agents(male_pos, obj_function, data)
female_pos, female_fitness = sort_agents(female_pos, obj_function, data)
#updating values of nuptial dance
d = d * delta
fl = fl * delta
#update final information
display(male_pos, male_fitness, agent_name)
if(male_fitness[0] > Leader_fitness):
Leader_agent = male_pos[0].copy()
Leader_fitness = male_fitness[0].copy()
convergence_curve['fitness'][iter_no] = Leader_fitness
convergence_curve['feature_count'][iter_no] = int(np.sum(Leader_agent))
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy, data)
male_pos, male_accuracy = sort_agents(male_pos, compute_accuracy, data)
print('\n================================================================================')
print(' Final Result ')
print('================================================================================\n')
print('Leader ' + agent_name + ' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name + ' Classification Accuracy : {}'.format(Leader_accuracy))
print('\n================================================================================\n')
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence curves
iters = np.arange(max_iter)+1
fig, axes = plt.subplots(2, 1)
fig.tight_layout(pad = 5)
fig.suptitle('Convergence Curves')
axes[0].set_title('Convergence of Fitness over Iterations')
axes[0].set_xlabel('Iteration')
axes[0].set_ylabel('Fitness')
axes[0].plot(iters, convergence_curve['fitness'])
axes[1].set_title('Convergence of Feature Count over Iterations')
axes[1].set_xlabel('Iteration')
axes[1].set_ylabel('Number of Selected Features')
axes[1].plot(iters, convergence_curve['feature_count'])
if(save_conv_graph):
plt.savefig('convergence_graph_'+ short_name + '.jpg')
plt.show()
# update attributes of solution
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_population = male_pos
solution.final_fitness = male_fitness
solution.final_accuracy = male_accuracy
solution.execution_time = exec_time
return solution
def PSO(num_agents,
max_iter,
train_data,
train_label,
obj_function=compute_fitness,
trans_func_shape='s',
save_conv_graph=False):
# Particle Swarm Optimizer
############################### Parameters ####################################
# #
# num_agents: number of particles #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# trans_function_shape: shape of the transfer function used #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
short_name = 'PSO'
agent_name = 'Particle'
train_data, train_label = np.array(train_data), np.array(train_label)
num_features = train_data.shape[1]
trans_function = get_trans_function(trans_func_shape)
# setting up the objectives
weight_acc = None
if (obj_function == compute_fitness):
weight_acc = float(
input('Weight for the classification accuracy [0-1]: '))
obj = (obj_function, weight_acc)
compute_accuracy = (
compute_fitness, 1
) # compute_accuracy is just compute_fitness with accuracy weight as 1
# initialize particles and Leader (the agent with the max fitness)
particles = initialize(num_agents, num_features)
fitness = np.zeros(num_agents)
accuracy = np.zeros(num_agents)
Leader_agent = np.zeros((1, num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
# initialize data class
data = Data()
val_size = float(
input('Enter the percentage of data wanted for valdiation [0, 100]: ')
) / 100
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(
train_data, train_label, stratify=train_label, test_size=val_size)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# rank initial particles
particles, fitness = sort_agents(particles, obj, data)
# start timer
start_time = time.time()
# initialize global and local best particles
globalBestParticle = [0 for i in range(num_features)]
globalBestFitness = float("-inf")
localBestParticle = [[0 for i in range(num_features)]
for j in range(num_agents)]
localBestFitness = [float("-inf") for i in range(num_agents)]
weight = 1.0
velocity = [[0.0 for i in range(num_features)] for j in range(num_agents)]
for iter_no in range(max_iter):
print(
'\n================================================================================'
)
print(' Iteration - {}'.format(iter_no + 1))
print(
'================================================================================\n'
)
# update weight
weight = 1.0 - (iter_no / max_iter)
# update the velocity
for i in range(num_agents):
for j in range(num_features):
velocity[i][j] = (weight * velocity[i][j])
r1, r2 = np.random.random(2)
velocity[i][j] = velocity[i][j] + (
r1 * (localBestParticle[i][j] - particles[i][j]))
velocity[i][j] = velocity[i][j] + (
r2 * (globalBestParticle[j] - particles[i][j]))
# updating position of particles
for i in range(num_agents):
for j in range(num_features):
trans_value = trans_function(velocity[i][j])
if (np.random.random() < trans_value):
particles[i][j] = 1
else:
particles[i][j] = 0
# updating fitness of particles
particles, fitness = sort_agents(particles, obj, data)
display(particles, fitness, agent_name)
# updating the global best and local best particles
for i in range(num_agents):
if fitness[i] > localBestFitness[i]:
localBestFitness[i] = fitness[i]
localBestParticle[i] = particles[i][:]
if fitness[i] > globalBestFitness:
globalBestFitness = fitness[i]
globalBestParticle = particles[i][:]
# update Leader (best agent)
if globalBestFitness > Leader_fitness:
Leader_agent = globalBestParticle.copy()
Leader_fitness = globalBestFitness.copy()
convergence_curve['fitness'][iter_no] = np.mean(fitness)
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy,
data)
particles, accuracy = sort_agents(particles, compute_accuracy, data)
print(
'\n================================================================================'
)
print(
' Final Result '
)
print(
'================================================================================\n'
)
print('Leader ' + agent_name +
' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name +
' Classification Accuracy : {}'.format(Leader_accuracy))
print(
'\n================================================================================\n'
)
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence graph
fig, axes = Conv_plot(convergence_curve)
if (save_conv_graph):
plt.savefig('convergence_graph_' + short_name + '.jpg')
plt.show()
# update attributes of solution
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_particles = particles
solution.final_fitness = fitness
solution.final_accuracy = accuracy
solution.execution_time = exec_time
return solution
def WOA(num_agents,
max_iter,
train_data,
train_label,
obj_function=compute_fitness,
trans_function_shape='s',
save_conv_graph=False):
# Whale Optimization Algorithm
############################### Parameters ####################################
# #
# num_agents: number of whales #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# trans_function_shape: shape of the transfer function used #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
short_name = 'WOA'
agent_name = 'Whale'
train_data, train_label = np.array(train_data), np.array(train_label)
num_features = train_data.shape[1]
cross_limit = 5
trans_function = get_trans_function(trans_function_shape)
# setting up the objectives
weight_acc = None
if (obj_function == compute_fitness):
weight_acc = float(
input('Weight for the classification accuracy [0-1]: '))
obj = (obj_function, weight_acc)
compute_accuracy = (
compute_fitness, 1
) # compute_accuracy is just compute_fitness with accuracy weight as 1
# initialize whales and Leader (the agent with the max fitness)
whales = initialize(num_agents, num_features)
fitness = np.zeros(num_agents)
accuracy = np.zeros(num_agents)
Leader_agent = np.zeros((1, num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
convergence_curve['feature_count'] = np.zeros(max_iter)
# format the data
data = Data()
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(
train_data, train_label, stratify=train_label, test_size=0.2)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# rank initial population
whales, fitness = sort_agents(whales, obj, data)
# start timer
start_time = time.time()
# main loop
for iter_no in range(max_iter):
print(
'\n================================================================================'
)
print(' Iteration - {}'.format(iter_no + 1))
print(
'================================================================================\n'
)
a = 2 - iter_no * (2 / max_iter) # a decreases linearly fron 2 to 0
# update the position of each whale
for i in range(num_agents):
# update the parameters
r = np.random.random() # r is a random number in [0, 1]
A = (2 * a * r) - a # Eq. (3)
C = 2 * r # Eq. (4)
l = -1 + (np.random.random() * 2
) # l is a random number in [-1, 1]
p = np.random.random() # p is a random number in [0, 1]
b = 1 # defines shape of the spiral
if p < 0.5:
# Shrinking Encircling mechanism
if abs(A) >= 1:
rand_agent_index = np.random.randint(0, num_agents)
rand_agent = whales[rand_agent_index, :]
mod_dist_rand_agent = abs(C * rand_agent - whales[i, :])
whales[i, :] = rand_agent - (A * mod_dist_rand_agent
) # Eq. (9)
else:
mod_dist_Leader = abs(C * Leader_agent - whales[i, :])
whales[i, :] = Leader_agent - (A * mod_dist_Leader
) # Eq. (2)
else:
# Spiral-Shaped Attack mechanism
dist_Leader = abs(Leader_agent - whales[i, :])
whales[i, :] = dist_Leader * np.exp(b * l) * np.cos(
l * 2 * np.pi) + Leader_agent
# Apply transformation function on the updated whale
for j in range(num_features):
trans_value = trans_function(whales[i, j])
if (np.random.random() < trans_value):
whales[i, j] = 1
else:
whales[i, j] = 0
# update final information
whales, fitness = sort_agents(whales, obj, data)
display(whales, fitness, agent_name)
if fitness[0] > Leader_fitness:
Leader_agent = whales[0].copy()
Leader_fitness = fitness[0].copy()
convergence_curve['fitness'][iter_no] = Leader_fitness
convergence_curve['feature_count'][iter_no] = int(np.sum(Leader_agent))
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy,
data)
whales, accuracy = sort_agents(whales, compute_accuracy, data)
print(
'\n================================================================================'
)
print(
' Final Result '
)
print(
'================================================================================\n'
)
print('Leader ' + agent_name +
' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name +
' Classification Accuracy : {}'.format(Leader_accuracy))
print(
'\n================================================================================\n'
)
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence curves
iters = np.arange(max_iter) + 1
fig, axes = plt.subplots(2, 1)
fig.tight_layout(pad=5)
fig.suptitle('Convergence Curves')
axes[0].set_title('Convergence of Fitness over Iterations')
axes[0].set_xlabel('Iteration')
axes[0].set_ylabel('Fitness')
axes[0].plot(iters, convergence_curve['fitness'])
axes[1].set_title('Convergence of Feature Count over Iterations')
axes[1].set_xlabel('Iteration')
axes[1].set_ylabel('Number of Selected Features')
axes[1].plot(iters, convergence_curve['feature_count'])
if (save_conv_graph):
plt.savefig('convergence_graph_' + short_name + '.jpg')
plt.show()
# update attributes of solution
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_population = whales
solution.final_fitness = fitness
solution.final_accuracy = accuracy
solution.execution_time = exec_time
return solution
def GWO(num_agents,
max_iter,
train_data,
train_label,
obj_function=compute_fitness,
trans_func_shape='s',
save_conv_graph=False):
# Grey Wolf Optimizer
############################### Parameters ####################################
# #
# num_agents: number of greywolves #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# trans_function_shape: shape of the transfer function used #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
short_name = 'GWO'
agent_name = 'Greywolf'
train_data, train_label = np.array(train_data), np.array(train_label)
num_features = train_data.shape[1]
trans_function = get_trans_function(trans_func_shape)
# setting up the objectives
weight_acc = None
if (obj_function == compute_fitness):
weight_acc = float(
input('Weight for the classification accuracy [0-1]: '))
obj = (obj_function, weight_acc)
compute_accuracy = (
compute_fitness, 1
) # compute_accuracy is just compute_fitness with accuracy weight as 1
# initialize greywolves and Leader (the agent with the max fitness)
greywolves = initialize(num_agents, num_features)
fitness = np.zeros(num_agents)
accuracy = np.zeros(num_agents)
Leader_agent = np.zeros((1, num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
convergence_curve['feature_count'] = np.zeros(max_iter)
# initialize data class
data = Data()
val_size = float(
input('Enter the percentage of data wanted for valdiation [0, 100]: ')
) / 100
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(
train_data, train_label, stratify=train_label, test_size=val_size)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# rank initial greywolves
greywolves, fitness = sort_agents(greywolves, obj, data)
# start timer
start_time = time.time()
# initialize the alpha, beta and delta grey wolves and their fitness
alpha, beta, delta = np.zeros((1, num_features)), np.zeros(
(1, num_features)), np.zeros((1, num_features))
alpha_fit, beta_fit, delta_fit = float("-inf"), float("-inf"), float(
"-inf")
for iter_no in range(max_iter):
print(
'\n================================================================================'
)
print(' Iteration - {}'.format(iter_no + 1))
print(
'================================================================================\n'
)
# update the alpha, beta and delta grey wolves
for i in range(num_agents):
# update alpha, beta, delta
if fitness[i] > alpha_fit:
delta_fit = beta_fit
delta = beta.copy()
beta_fit = alpha_fit
beta = alpha.copy()
alpha_fit = fitness[i]
alpha = greywolves[i, :].copy()
# update beta, delta
elif fitness[i] > beta_fit:
delta_fit = beta_fit
delta = beta.copy()
beta_fit = fitness[i]
beta = greywolves[i, :].copy()
# update delta
elif fitness[i] > delta_fit:
delta_fit = fitness[i]
delta = greywolves[i, :].copy()
# a decreases linearly fron 2 to 0
a = 2 - iter_no * ((2) / max_iter)
for i in range(num_agents):
for j in range(num_features):
# calculate distance between alpha and current agent
r1 = np.random.random() # r1 is a random number in [0,1]
r2 = np.random.random() # r2 is a random number in [0,1]
A1 = (2 * a * r1) - a # calculate A1
C1 = 2 * r2 # calculate C1
D_alpha = abs(C1 * alpha[j] -
greywolves[i, j]) # find distance from alpha
X1 = alpha[j] - (A1 * D_alpha) # Eq. (3.6)
# calculate distance between beta and current agent
r1 = np.random.random() # r1 is a random number in [0,1]
r2 = np.random.random() # r2 is a random number in [0,1]
A2 = (2 * a * r1) - a # calculate A2
C2 = 2 * r2 # calculate C2
D_beta = abs(C2 * beta[j] -
greywolves[i, j]) # find distance from beta
X2 = beta[j] - (A2 * D_beta) # Eq. (3.6)
# calculate distance between delta and current agent
r1 = np.random.random() # r1 is a random number in [0,1]
r2 = np.random.random() # r2 is a random number in [0,1]
A3 = (2 * a * r1) - a # calculate A3
C3 = 2 * r2 # calculate C3
D_delta = abs(C3 * delta[j] -
greywolves[i, j]) # find distance from delta
X3 = delta[j] - A3 * D_delta # Eq. (3.6)
# update the position of current agent
greywolves[i, j] = (X1 + X2 + X3) / 3 # Eq. (3.7)
# Apply transformation function on the updated greywolf
for j in range(num_features):
trans_value = trans_function(greywolves[i, j])
if (np.random.random() < trans_value):
greywolves[i, j] = 1
else:
greywolves[i, j] = 0
# update final information
greywolves, fitness = sort_agents(greywolves, obj, data)
display(greywolves, fitness, agent_name)
# update Leader (best agent)
if fitness[0] > Leader_fitness:
Leader_agent = greywolves[0].copy()
Leader_fitness = fitness[0].copy()
if alpha_fit > Leader_fitness:
Leader_fitness = alpha_fit
Leader_agent = alpha.copy()
convergence_curve['fitness'][iter_no] = Leader_fitness
convergence_curve['feature_count'][iter_no] = int(np.sum(Leader_agent))
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy,
data)
greywolves, accuracy = sort_agents(greywolves, compute_accuracy, data)
print(
'\n================================================================================'
)
print(
' Final Result '
)
print(
'================================================================================\n'
)
print('Leader ' + agent_name +
' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name +
' Classification Accuracy : {}'.format(Leader_accuracy))
print(
'\n================================================================================\n'
)
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence curves
iters = np.arange(max_iter) + 1
fig, axes = plt.subplots(2, 1)
fig.tight_layout(pad=5)
fig.suptitle('Convergence Curves')
axes[0].set_title('Convergence of Fitness over Iterations')
axes[0].set_xlabel('Iteration')
axes[0].set_ylabel('Fitness')
axes[0].plot(iters, convergence_curve['fitness'])
axes[1].set_title('Convergence of Feature Count over Iterations')
axes[1].set_xlabel('Iteration')
axes[1].set_ylabel('Number of Selected Features')
axes[1].plot(iters, convergence_curve['feature_count'])
if (save_conv_graph):
plt.savefig('convergence_graph_' + short_name + '.jpg')
plt.show()
# update attributes of solution
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_greywolves = greywolves
solution.final_fitness = fitness
solution.final_accuracy = accuracy
solution.execution_time = exec_time
return solution
def CS (num_agents, max_iter, train_data, train_label, obj_function=compute_fitness, trans_function_shape='s', save_conv_graph=False):
# Cuckoo Search Algorithm
############################### Parameters ####################################
# #
# num_agents: number of agents #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# trans_function_shape: shape of the transfer function used #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
short_name = 'CS'
agent_name = 'Agent'
num_features = train_data.shape[1]
trans_function = get_trans_function(trans_function_shape)
# setting up the objectives
weight_acc = None
if(obj_function==compute_fitness):
weight_acc = float(input('Weight for the classification accuracy [0-1]: '))
obj = (obj_function, weight_acc)
compute_accuracy = (compute_fitness, 1) # compute_accuracy is just compute_fitness with accuracy weight as 1
# initializing cuckoo and host nests
levy_flight = np.random.uniform(low=-2, high=2, size=(num_features))
cuckoo = np.random.randint(low=0, high=2, size=(num_features))
nest = initialize(num_agents, num_features)
fitness = np.zeros(num_agents)
nest_accuracy = np.zeros(num_agents)
cuckoo_fitness = float("-inf")
Leader_agent = np.zeros((num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
p_a=0.25 # fraction of nests to be replaced
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
# initialize data class
data = Data()
val_size = float(input('Enter the percentage of data wanted for valdiation [0, 100]: '))/100
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(train_data, train_label, stratify=train_label, test_size=val_size)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# rank initial nests
nest, fitness = sort_agents(nest, obj, data)
cuckoo,cuckoo_fitness = sort_agents(cuckoo,obj,data)
# start timer
start_time = time.time()
# main loop
for iter_no in range(max_iter):
print('\n================================================================================')
print(' Iteration - {}'.format(iter_no+1))
print('================================================================================\n')
# updating leader nest
if fitness[0] > Leader_fitness:
Leader_agent = nest[0].copy()
Leader_fitness = fitness[0]
# get new cuckoo
levy_flight = get_cuckoo(levy_flight)
for j in range(num_features):
if trans_function(levy_flight[j]) > np.random.random():
cuckoo[j]=1
else:
cuckoo[j]=0
cuckoo,cuckoo_fitness = sort_agents(cuckoo,obj,data)
# check if a nest needs to be replaced
j = np.random.randint(0,num_agents)
if cuckoo_fitness > fitness[j]:
nest[j] = cuckoo.copy()
fitness[j] = cuckoo_fitness
nest, fitness = sort_agents(nest, obj, data)
# eliminate worse nests and generate new ones
nest = replace_worst(nest, p_a)
nest, fitness = sort_agents(nest, obj, data)
# update final information
display(nest, fitness, agent_name)
if fitness[0]>Leader_fitness:
Leader_agent = nest[0].copy()
Leader_fitness = fitness[0].copy()
convergence_curve['fitness'][iter_no] = np.mean(fitness)
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy, data)
nest, nest_accuracy = sort_agents(nest, compute_accuracy, data)
print('\n================================================================================')
print(' Final Result ')
print('================================================================================\n')
print('Leader ' + agent_name + ' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name + ' Classification Accuracy : {}'.format(Leader_accuracy))
print('\n================================================================================\n')
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence graph
fig, axes = Conv_plot(convergence_curve)
if(save_conv_graph):
plt.savefig('convergence_graph_'+ short_name + '.jpg')
plt.show()
# update attributes of solution
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_population = nest
solution.final_fitness = fitness
solution.final_accuracy = nest_accuracy
solution.execution_time = exec_time
return solution
def CS(num_agents,
max_iter,
train_data,
train_label,
obj_function=compute_fitness,
trans_function_shape='s',
save_conv_graph=False):
# Cuckoo Search Algorithm
############################### Parameters ####################################
# #
# num_agents: number of agents #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# trans_function_shape: shape of the transfer function used #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
short_name = 'CS'
agent_name = 'Agent'
num_features = train_data.shape[1]
trans_function = get_trans_function(trans_function_shape)
# initializing cuckoo and host nests
levy_flight = np.random.uniform(low=-2, high=2, size=(num_features))
cuckoo = np.random.randint(low=0, high=2, size=(num_features))
nest = initialize(num_agents, num_features)
nest_fitness = np.zeros(num_agents)
nest_accuracy = np.zeros(num_agents)
cuckoo_fitness = float("-inf")
Leader_agent = np.zeros((num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
p_a = 0.25 # fraction of nests to be replaced
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
convergence_curve['feature_count'] = np.zeros(max_iter)
data = Data()
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(
train_data, train_label, stratify=train_label, test_size=0.2)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# rank initial nests
nest, nest_fitness = sort_agents(nest, obj_function, data)
# start timer
start_time = time.time()
# main loop
for iter_no in range(max_iter):
print(
'\n================================================================================'
)
print(' Iteration - {}'.format(iter_no + 1))
print(
'================================================================================\n'
)
# updating leader nest
if nest_fitness[0] > Leader_fitness:
Leader_agent = nest[0].copy()
Leader_fitness = nest_fitness[0]
# get new cuckoo
levy_flight = get_cuckoo(levy_flight)
for j in range(num_features):
if trans_function(levy_flight[j]) > np.random.random():
cuckoo[j] = 1
else:
cuckoo[j] = 0
# check if a nest needs to be replaced
j = np.random.randint(0, num_agents)
if cuckoo_fitness > nest_fitness[j]:
nest[j] = cuckoo.copy()
nest_fitness[j] = cuckoo_fitness
nest, nest_fitness = sort_agents(nest, obj_function, data)
# eliminate worse nests and generate new ones
nest = replace_worst(nest, p_a)
nest, nest_fitness = sort_agents(nest, obj_function, data)
# update final information
display(nest, nest_fitness, agent_name)
if nest_fitness[0] > Leader_fitness:
Leader_agent = nest[0].copy()
Leader_fitness = nest_fitness[0].copy()
convergence_curve['fitness'][iter_no] = Leader_fitness
convergence_curve['feature_count'][iter_no] = int(np.sum(Leader_agent))
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy,
data)
nest, nest_accuracy = sort_agents(nest, compute_accuracy, data)
print(
'\n================================================================================'
)
print(
' Final Result '
)
print(
'================================================================================\n'
)
print('Leader ' + agent_name +
' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name +
' Classification Accuracy : {}'.format(Leader_accuracy))
print(
'\n================================================================================\n'
)
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence curves
iters = np.arange(max_iter) + 1
fig, axes = plt.subplots(2, 1)
fig.tight_layout(pad=5)
fig.suptitle('Convergence Curves')
axes[0].set_title('Convergence of Fitness over Iterations')
axes[0].set_xlabel('Iteration')
axes[0].set_ylabel('Fitness')
axes[0].plot(iters, convergence_curve['fitness'])
axes[1].set_title('Convergence of Feature Count over Iterations')
axes[1].set_xlabel('Iteration')
axes[1].set_ylabel('Number of Selected Features')
axes[1].plot(iters, convergence_curve['feature_count'])
if (save_conv_graph):
plt.savefig('convergence_graph_' + short_name + '.jpg')
plt.show()
# update attributes of solution
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_population = nest
solution.final_fitness = nest_fitness
solution.final_accuracy = nest_accuracy
solution.execution_time = exec_time
return solution
def GA(num_agents,
max_iter,
train_data,
train_label,
obj_function=compute_fitness,
prob_cross=0.4,
prob_mut=0.3,
save_conv_graph=False,
seed=0):
# Genetic Algorithm
############################### Parameters ####################################
# #
# num_agents: number of chromosomes #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# prob_cross: probability of crossover #
# prob_mut: probability of mutation #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
short_name = 'GA'
agent_name = 'Chromosome'
train_data, train_label = np.array(train_data), np.array(train_label)
num_features = train_data.shape[1]
cross_limit = 5
np.random.seed(seed)
# setting up the objectives
weight_acc = None
if (obj_function == compute_fitness):
weight_acc = float(
input('Weight for the classification accuracy [0-1]: '))
obj = (obj_function, weight_acc)
compute_accuracy = (
compute_fitness, 1
) # compute_accuracy is just compute_fitness with accuracy weight as 1
# initialize chromosomes and Leader (the agent with the max fitness)
chromosomes = initialize(num_agents, num_features)
fitness = np.zeros(num_agents)
accuracy = np.zeros(num_agents)
Leader_agent = np.zeros((1, num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
# initialize data class
data = Data()
val_size = float(
input('Enter the percentage of data wanted for valdiation [0, 100]: ')
) / 100
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(
train_data, train_label, stratify=train_label, test_size=val_size)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# rank initial population
chromosomes, fitness = sort_agents(chromosomes, obj, data)
# start timer
start_time = time.time()
# main loop
for iter_no in range(max_iter):
print(
'\n================================================================================'
)
print(' Iteration - {}'.format(iter_no + 1))
print(
'================================================================================\n'
)
# perform crossover, mutation and replacement
chromosomes, fitness = cross_mut(chromosomes, fitness, obj, data,
prob_cross, cross_limit, prob_mut)
# update final information
chromosomes, fitness = sort_agents(chromosomes, obj, data, fitness)
display(chromosomes, fitness, agent_name)
if fitness[0] > Leader_fitness:
Leader_agent = chromosomes[0].copy()
Leader_fitness = fitness[0].copy()
convergence_curve['fitness'][iter_no] = np.mean(fitness)
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy,
data)
chromosomes, accuracy = sort_agents(chromosomes, compute_accuracy, data)
print(
'\n================================================================================'
)
print(
' Final Result '
)
print(
'================================================================================\n'
)
print('Leader ' + agent_name +
' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name +
' Classification Accuracy : {}'.format(Leader_accuracy))
print(
'\n================================================================================\n'
)
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence graph
fig, axes = Conv_plot(convergence_curve)
if (save_conv_graph):
plt.savefig('convergence_graph_' + short_name + '.jpg')
plt.show()
# update attributes of solution
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_population = chromosomes
solution.final_fitness = fitness
solution.final_accuracy = accuracy
solution.execution_time = exec_time
return solution
def GA(num_agents, max_iter, train_data, train_label, obj_function=compute_fitness, prob_cross=0.4, prob_mut=0.3, save_conv_graph=False):
# Genetic Algorithm
############################### Parameters ####################################
# #
# num_agents: number of chromosomes #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# prob_cross: probability of crossover #
# prob_mut: probability of mutation #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
short_name = 'GA'
agent_name = 'Chromosome'
train_data, train_label = np.array(train_data), np.array(train_label)
num_features = train_data.shape[1]
cross_limit = 5
# initialize chromosomes and Leader (the agent with the max fitness)
chromosomes = initialize(num_agents, num_features)
fitness = np.zeros(num_agents)
accuracy = np.zeros(num_agents)
Leader_agent = np.zeros((1, num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
convergence_curve['feature_count'] = np.zeros(max_iter)
# format the data
data = Data()
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(train_data, train_label, stratify=train_label, test_size=0.2)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# rank initial population
chromosomes, fitness = sort_agents(chromosomes, obj_function, data)
# start timer
start_time = time.time()
# main loop
for iter_no in range(max_iter):
print('\n================================================================================')
print(' Iteration - {}'.format(iter_no+1))
print('================================================================================\n')
# perform crossover, mutation and replacement
cross_mut(chromosomes, fitness, obj_function, data, prob_cross, cross_limit, prob_mut)
# update final information
chromosomes, fitness = sort_agents(chromosomes, obj_function, data)
display(chromosomes, fitness, agent_name)
if fitness[0]>Leader_fitness:
Leader_agent = chromosomes[0].copy()
Leader_fitness = fitness[0].copy()
convergence_curve['fitness'][iter_no] = Leader_fitness
convergence_curve['feature_count'][iter_no] = int(np.sum(Leader_agent))
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy, data)
chromosomes, accuracy = sort_agents(chromosomes, compute_accuracy, data)
print('\n================================================================================')
print(' Final Result ')
print('================================================================================\n')
print('Leader ' + agent_name + ' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name + ' Classification Accuracy : {}'.format(Leader_accuracy))
print('\n================================================================================\n')
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence curves
iters = np.arange(max_iter)+1
fig, axes = plt.subplots(2, 1)
fig.tight_layout(pad = 5)
fig.suptitle('Convergence Curves')
axes[0].set_title('Convergence of Fitness over Iterations')
axes[0].set_xlabel('Iteration')
axes[0].set_ylabel('Fitness')
axes[0].plot(iters, convergence_curve['fitness'])
axes[1].set_title('Convergence of Feature Count over Iterations')
axes[1].set_xlabel('Iteration')
axes[1].set_ylabel('Number of Selected Features')
axes[1].plot(iters, convergence_curve['feature_count'])
if(save_conv_graph):
plt.savefig('convergence_graph_'+ short_name + '.jpg')
plt.show()
# update attributes of solution
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_population = chromosomes
solution.final_fitness = fitness
solution.final_accuracy = accuracy
solution.execution_time = exec_time
return solution
def Name_of_the_wrapper(num_agents,
max_iter,
train_data,
train_label,
obj_function=compute_fitness,
trans_func_shape='s',
save_conv_graph=False):
# Name of the optimizer
############################### Parameters ####################################
# #
# num_agents: number of agents #
# max_iter: maximum number of generations #
# train_data: training samples of data #
# train_label: class labels for the training samples #
# obj_function: the function to maximize while doing feature selection #
# trans_function_shape: shape of the transfer function used #
# save_conv_graph: boolean value for saving convergence graph #
# #
###############################################################################
short_name = ''
agent_name = ''
train_data, train_label = np.array(train_data), np.array(train_label)
num_features = train_data.shape[1]
trans_function = get_trans_function(trans_func_shape)
# setting up the objectives
weight_acc = None
if (obj_function == compute_fitness):
weight_acc = float(
input('Weight for the classification accuracy [0-1]: '))
obj = (obj_function, weight_acc)
compute_accuracy = (
compute_fitness, 1
) # compute_accuracy is just compute_fitness with accuracy weight as 1
# initialize agents and Leader (the agent with the max fitness)
agents = initialize(num_agents, num_features)
fitness = np.zeros(num_agents)
accuracy = np.zeros(num_agents)
Leader_agent = np.zeros((1, num_features))
Leader_fitness = float("-inf")
Leader_accuracy = float("-inf")
# initialize convergence curves
convergence_curve = {}
convergence_curve['fitness'] = np.zeros(max_iter)
convergence_curve['feature_count'] = np.zeros(max_iter)
# format the data
data = Data()
data.train_X, data.val_X, data.train_Y, data.val_Y = train_test_split(
train_data, train_label, stratify=train_label, test_size=0.2)
# create a solution object
solution = Solution()
solution.num_agents = num_agents
solution.max_iter = max_iter
solution.num_features = num_features
solution.obj_function = obj_function
# rank initial agents
agents, fitness = sort_agents(agents, obj, data)
# start timer
start_time = time.time()
for iter_no in range(max_iter):
print(
'\n================================================================================'
)
print(' Iteration - {}'.format(iter_no + 1))
print(
'================================================================================\n'
)
################ write your main position update code here ################
###########################################################################
# update final information
agents, fitness = sort_agents(agents, obj, data)
display(agents, fitness, agent_name)
# update Leader (best agent)
if fitness[0] > Leader_fitness:
Leader_agent = agents[0].copy()
Leader_fitness = fitness[0].copy()
convergence_curve['fitness'][iter_no] = Leader_fitness
convergence_curve['feature_count'][iter_no] = int(np.sum(Leader_agent))
# compute final accuracy
Leader_agent, Leader_accuracy = sort_agents(Leader_agent, compute_accuracy,
data)
agents, accuracy = sort_agents(agents, compute_accuracy, data)
print(
'\n================================================================================'
)
print(
' Final Result '
)
print(
'================================================================================\n'
)
print('Leader ' + agent_name +
' Dimension : {}'.format(int(np.sum(Leader_agent))))
print('Leader ' + agent_name + ' Fitness : {}'.format(Leader_fitness))
print('Leader ' + agent_name +
' Classification Accuracy : {}'.format(Leader_accuracy))
print(
'\n================================================================================\n'
)
# stop timer
end_time = time.time()
exec_time = end_time - start_time
# plot convergence graph
fig, axes = Conv_plot(convergence_curve)
if (save_conv_graph):
plt.savefig('convergence_graph_' + short_name + '.jpg')
plt.show()
# update attributes of solution
solution.best_agent = Leader_agent
solution.best_fitness = Leader_fitness
solution.best_accuracy = Leader_accuracy
solution.convergence_curve = convergence_curve
solution.final_agents = agents
solution.final_fitness = fitness
solution.final_accuracy = accuracy
solution.execution_time = exec_time
return solution
