def test_random_features_generator(sample_size=10,
compute_sample_statistics=True):
config = Config()
config.num_true_features = 2
config.num_obs_features = 2
config.max_num_features = 20000
task = RandomFeatures(config)
print("The value of theta is:\n{0}".format(task.theta))
print("The norm of theta is: {0}".format(np.linalg.norm(task.theta)))
for i in range(sample_size):
target, observable_features, best_approximation = task.sample_observation(
noisy=False)
print("The features are: {0}\tThe target is:{1}".format(
observable_features, target))
if compute_sample_statistics:
num_samples = 100000
samples = np.zeros(num_samples)
for i in range(num_samples):
target, _, _ = task.sample_observation(noisy=False)
samples[i] += target
# The sample average and sample variance of the target should be 0 and 1, respectively.
print("The sample average of the target is: {:.2f}".format(
np.average(samples)))
print("The sample variance of the target is: {:.2f}".format(
np.var(samples)))
def test_function_approximator(num_features=20,
initial_features=20,
num_iterations=10000,
chkpt=100,
plot_mse=True,
noisy=True,
add_features=False,
add_true_features=True,
feature_add_interval=100,
mixed_features=False):
from src.step_size_methods import SGD
config = Config()
# task setup
config.num_true_features = num_features
config.num_obs_features = initial_features # same as function approximator
config.max_num_features = 20000 # same as function approximator
task = RandomFeatures(config)
# function approximator setup
approximator = LinearFunctionApproximator(config)
# optimizer setup
config.parameter_size = initial_features
config.alpha = 0.001
optimizer = SGD(config)
# for plotting
mse_per_chpt = np.zeros(num_iterations // chkpt, dtype=np.float64)
mse = 0
current_chpt = 0
# training loop
for i in range(num_iterations):
target, observable_features, best_approximation = task.sample_observation(
noisy=noisy)
prediction = approximator.get_prediction(observable_features)
error = target - prediction
_, _, new_weights = optimizer.update_weight_vector(
error, observable_features, approximator.get_weight_vector())
approximator.update_weight_vector(new_weights)
squared_loss = np.square(error)
mse += squared_loss / chkpt
if (i + 1) % chkpt == 0:
# reporting and saving
print("Iteration number: {0}".format(i + 1))
print("\tTarget: {0:.4f}".format(target))
print("\tPrediction: {0:.4f}".format(prediction))
print("\tMean Squared Error: {0:.4f}".format(mse))
mse_per_chpt[current_chpt] += mse
mse *= 0
current_chpt += 1
if add_features and (i + 1) % feature_add_interval == 0:
task.add_new_feature(k=1, true_feature=add_true_features)
approximator.increase_num_features(k=1)
optimizer.increase_size(k=1)
if mixed_features:
add_true_features = not add_true_features
if plot_mse:
# plots
import matplotlib.pyplot as plt
x_axis = np.arange(num_iterations // chkpt)
plt.plot(x_axis, mse_per_chpt)
plt.show()
plt.close()