def make_viz():
"""Produce visualization."""
# Set plot settings
ylim = [-5, 5]
gpv.update_rc_params()
# Define model params
gp1_hypers = {'ls': 1., 'alpha': 1.5, 'sigma': 1e-1}
gp2_hypers = {'ls': 3., 'alpha': 1., 'sigma': 1e-1}
# Define domain
x_min = -10.
x_max = 10.
domain = RealDomain({'min_max': [(x_min, x_max)]})
# Drawn one sample path from gp1
gp1 = SimpleGp(gp1_hypers)
sample_path = gput.get_sample_path(gp1, domain)
n_obs_list = [3, 5, 10, 20, 30, 60]
# Make full data
data_full = gput.get_data_from_sample_path(
sample_path, gp1.params.sigma*4.0, 100
)
# Define set of domain points
dom_pt_list = list(np.linspace(x_min, x_max, 200))
for n_obs in n_obs_list:
data = get_data_subset(data_full, n_obs)
# Define gp2 (posterior)
gp2 = SimpleGp(gp2_hypers)
gp2.set_data(data)
# Define gp3 (modified prior)
gp3_hypers = copy.deepcopy(gp2_hypers)
gp3_hypers['alpha'] += 0.01
gp3 = SimpleGp(gp3_hypers)
lb_list, ub_list = get_lb_ub_lists(dom_pt_list, gp2, gp3, data, False)
# Various plotting
plt.figure()
# Plot C_t bounds
plt.plot(dom_pt_list, lb_list, 'g--')
plt.plot(dom_pt_list, ub_list, 'g--')
plt.fill_between(dom_pt_list, lb_list, ub_list, color='wheat', alpha=0.5)
# Plot GP posterior
save_str = 'viz_' + str(n_obs)
gpv.visualize_gp_and_sample_path(gp2, domain, data, sample_path,
show_sp=False, ylim=ylim,
save_str=save_str)
plt.close()
def make_viz():
"""Produce visualization."""
# Set plot settings
ylim = [-5, 5]
gpv.update_rc_params()
# Define model params
gp1_hypers = {'ls': 1., 'alpha': 1.5, 'sigma': 1e-1}
gp2_hypers = {'ls': 3., 'alpha': 1., 'sigma': 1e-1}
# Define domain
x_min = -10.
x_max = 10.
domain = RealDomain({'min_max': [(x_min, x_max)]})
# Define gp1 and gp2
gp1 = SimpleGp(gp1_hypers)
gp2 = SimpleGp(gp2_hypers)
# Plot GP priors
data = Namespace(X=[], y=np.array([]))
gpv.visualize_gp(gp1, domain, data, std_mult=2, ylim=ylim, save_str='viz_prior_true')
plt.close()
gpv.visualize_gp(gp2, domain, data, std_mult=2, ylim=ylim, save_str='viz_prior')
plt.close()
def run_gplcb():
"""Run LCB algorithm with Gaussian process (GP-LCB)."""
# Set plot settings
ylim = [-5, 5]
gpv.update_rc_params()
# Define model params
gp1_hypers = {'ls': 1., 'alpha': 1.5, 'sigma': 1e-1}
gp2_hypers = {'ls': 3., 'alpha': 1., 'sigma': 1e-1}
# Define domain
x_min = -10.
x_max = 10.
domain = RealDomain({'min_max': [(x_min, x_max)]})
# Drawn one sample path from gp1
gp1 = SimpleGp(gp1_hypers)
sample_path_nonoise = gput.get_sample_path(gp1, domain)
# Convert to noisy sample_path observations
sample_path = gput.get_noisy_sample_path(sample_path_nonoise,
gp1.params.sigma)
# Setup BO: initialize data
data = Namespace(X=[], y=np.array([]))
init_x = np.random.choice(sample_path.x, 5)[3]
data = query_function_update_data(init_x, data, sample_path)
# Define set of domain points
dom_pt_list = list(sample_path.x) # NOTE: confirm if correct
n_iter = 30
print('Finished iter: ', end='')
for i in range(n_iter):
# Define gp2 (posterior)
gp2 = SimpleGp(gp2_hypers)
gp2.set_data(data)
# Compute lb_list
test_pts = [np.array([dom_pt]) for dom_pt in dom_pt_list]
mean_list, std_list = gp2.get_gp_post_mu_cov(test_pts, full_cov=False)
conf_mean = np.array(mean_list)
std_mult = 3.
lb_list = conf_mean - std_mult * np.array(
std_list) # technically is np ndarray
ub_list = conf_mean + std_mult * np.array(
std_list) # technically is np ndarray
# Compute next_query_point
min_idx = np.nanargmin(lb_list)
next_query_point = dom_pt_list[min_idx]
# Various plotting
plt.figure()
# Plot C_t bounds
plt.plot(dom_pt_list, mean_list, 'k--', lw=2)
plt.plot(dom_pt_list, lb_list, 'k--')
plt.plot(dom_pt_list, ub_list, 'k--')
plt.fill_between(dom_pt_list,
lb_list,
ub_list,
color='lightsteelblue',
alpha=0.5)
# Plot GP posterior
save_str = 'viz_' + str(i)
gpv.visualize_sample_path_and_data(sample_path_nonoise,
data,
ylim=ylim,
save_str=save_str)
plt.close()
# update data
data = query_function_update_data(next_query_point, data, sample_path)
print('{}, '.format(i), end='')
print('Data:')
print(data)
def run_cslcb():
"""Run LCB algorithm with confidence sequence (CS-LCB)."""
# Set plot settings
ylim = [-5, 5]
gpv.update_rc_params()
# Define model params
gp1_hypers = {'ls': 1., 'alpha': 1.5, 'sigma': 1e-1}
gp2_hypers = {'ls': 1., 'alpha': 1.5, 'sigma': 1e-1}
# Define domain
x_min = -10.
x_max = 10.
domain = RealDomain({'min_max': [(x_min, x_max)]})
# Drawn one sample path from gp1
gp1 = SimpleGp(gp1_hypers)
sample_path_nonoise = gput.get_sample_path(gp1, domain)
# Convert to noisy sample_path observations
sample_path = gput.get_noisy_sample_path(sample_path_nonoise,
gp1.params.sigma)
# Setup BO: initialize data
data = Namespace(X=[], y=np.array([]))
init_x = np.random.choice(sample_path.x, 5)[2]
data = query_function_update_data(init_x, data, sample_path)
# Define set of domain points
dom_pt_list = list(sample_path.x) # NOTE: confirm if correct
n_iter = 30
print('Finished iter: ', end='')
for i in range(n_iter):
# Define gp2 (posterior)
gp2 = SimpleGp(gp2_hypers)
gp2.set_data(data)
# Define gp3 (modified prior)
gp3_hypers = copy.deepcopy(gp2_hypers)
gp3_hypers['alpha'] += 0.01
gp3 = SimpleGp(gp3_hypers)
lb_list, ub_list = get_lb_ub_lists(dom_pt_list, gp2, gp3, data, False)
min_idx = np.nanargmin(lb_list)
next_query_point = dom_pt_list[min_idx]
# Various plotting
plt.figure()
# Plot C_t bounds
plt.plot(dom_pt_list, lb_list, 'g--')
plt.plot(dom_pt_list, ub_list, 'g--')
plt.fill_between(dom_pt_list,
lb_list,
ub_list,
color='wheat',
alpha=0.5)
# Plot GP posterior
save_str = 'viz_' + str(i)
gpv.visualize_sample_path_and_data(sample_path_nonoise,
data,
ylim=ylim,
save_str=save_str)
plt.close()
# update data
data = query_function_update_data(next_query_point, data, sample_path)
print('{}, '.format(i), end='')
print('Data:')
print(data)