def imperfect_features_test():
from src.env.Amatrix_task import Amatrix
n = 10
m = 2
env = Amatrix(n, m)
features = env.get_approx_A() # first m features
weights = np.zeros(m)
config = Config()
config.parameter_size = m
config.theta = 0.001
config.init_beta = 0
idbd = SIDBD(config)
sample_size = 50000
for i in range(sample_size):
rand_row = np.random.randint(n)
target = env.sample_target(rand_row, noisy=True)
pred_features = features[rand_row, :]
prediction = np.dot(pred_features, weights)
error = target - prediction
gradient, new_stepsize, new_weight_vector = idbd.update_weight_vector(
error, pred_features, weights)
weights = new_weight_vector
print("Sample number: {0}".format(i + 1))
print("\tPrediction error:{0}".format(error))
print("Theta star:\n{0}".format(env.theta_star))
print("Estimated theta:\n{0}".format(weights))
def perfect_features_test():
from src.env.Amatrix_task import Amatrix
n = 3
m = 2
env = Amatrix(n, m)
features = env.Amatrix # perfect features
weights = np.zeros(n)
config = Config()
config.parameter_size = n
config.theta = 0.1
config.init_beta = np.log(0.001)
idbd = SIDBD(config)
sample_size = 100000
for i in range(sample_size):
rand_row = np.random.randint(n)
target = env.sample_target(rand_row, noisy=True)
pred_features = features[rand_row, :]
prediction = np.dot(pred_features, weights)
error = target - prediction
gradient, new_stepsize, new_weight_vector = idbd.update_weight_vector(
error, pred_features, weights)
weights = new_weight_vector
if (i + 1) % 10000 == 0:
print("Sample number: {0}".format(i + 1))
print("\tPrediction error:{0}".format(error))
print("Theta star:\n{0}".format(env.theta_star))
print("Estimated theta:\n{0}".format(weights))
difference = np.sqrt(np.sum(np.square(env.theta_star - weights)))
print("L2 norm of difference:\n{0}".format(difference))
def adding_bad_features_test():
from src.env.Amatrix_task import Amatrix
n = 10
m = 5
env = Amatrix(n, m)
features = env.get_approx_A() # first m features
weights = np.zeros(m)
config = Config()
config.parameter_size = m
config.theta = 0.1
config.init_beta = np.log(0.0001)
idbd = SIDBD(config)
sample_size = 50000
additional_features = 30
for k in range(additional_features + 1):
print("Number of features in the representation: {0}".format(
idbd.parameter_size))
for i in range(sample_size):
rand_row = np.random.randint(n)
target = env.sample_target(rand_row, noisy=True)
pred_features = features[rand_row, :]
prediction = np.dot(pred_features, weights)
error = target - prediction
gradient, new_stepsize, new_weight_vector = idbd.update_weight_vector(
error, pred_features, weights)
weights = new_weight_vector
if ((i + 1) % 25000) == 0:
print("\tSample number: {0}".format(i + 1))
print("\t\tPrediction error: {0}".format(error))
print("Theta star:\n{0}".format(env.theta_star))
print("Estimated theta:\n{0}".format(weights))
if k < additional_features:
print("Adding new feature...")
new_feature = env.get_new_bad_features(1)
features = np.hstack((features, new_feature))
idbd.increase_size(1)
new_weights = np.zeros(m + 1)
new_weights[:m] = weights
m += 1
weights = new_weights