def _cov_mat(self, x1, x2=None, noise=True):
"""
Covariance matrix for preference data using the kernel function.
:param x1: datapoints for which to compute covariance matrix
:param x2: if None, covariance matrix will be square for the input x1
if not None, covariance will be between x1 (rows) and x2 (cols_
:param noise: whether to add noise to the diagonal of the covariance matrix
:return:
"""
if x2 is None:
x2 = x1
else: # if x1 != x2 we don't add noise!
noise = False
x1 = utils_data.format_data(x1, self.num_objectives)
x2 = utils_data.format_data(x2, self.num_objectives)
cov_mat = self._kernel(np.repeat(x1, x2.shape[0], axis=0),
np.tile(x2, (x1.shape[0], 1)))
cov_mat = cov_mat.reshape((x1.shape[0], x2.shape[0]))
if noise:
cov_mat += self.std_noise**2 * np.eye(cov_mat.shape[0])
return cov_mat
def _kernel(self, x1, x2):
"""
Squared exponential kernel function
:param x1:
:param x2:
:return:
"""
x1 = utils_data.format_data(x1, self.num_objectives)
x2 = utils_data.format_data(x2, self.num_objectives)
k = 0.8**2 * np.exp(-(1. / (2. * (self.kernel_width**2))) *
np.linalg.norm(x1 - x2, axis=1)**2)
return k
def get_predictive_params(self, x_new, pointwise):
"""
Returns the predictive parameters (mean, variance) of the Gaussian distribution
at the given datapoints
:param x_new: the points for which we want the predictive params
:param pointwise: whether we want pointwise variance or the entire covariance matrix
:return:
"""
# bring input points into right shape
x_new = utils_data.format_data(x_new, self.num_objectives)
# if we don't have any data yet, use prior GP to make predictions
if self.datapoints is None or self.utility_vals is None:
pred_mean, pred_var = self._evaluate_prior(x_new)
# otherwise compute predictive mean and covariance
else:
cov_xnew_x = self._cov_mat(x_new, self.datapoints, noise=False)
cov_x_xnew = self._cov_mat(self.datapoints, x_new, noise=False)
cov_xnew = self._cov_mat(x_new, noise=False)
pred_mean = self.prior_mean(x_new) + np.dot(
np.dot(cov_xnew_x, self.cov_mat_inv),
(self.utility_vals - self.prior_mean(self.datapoints)))
pred_var = cov_xnew - np.dot(
np.dot(cov_xnew_x, self.pred_cov_factor), cov_x_xnew)
if pointwise:
pred_var = pred_var.diagonal()
return pred_mean, pred_var
def get_predictive_params(self, x_new, pointwise):
# make sure data is in right shape
x_new = utils_data.format_data(x_new, self.num_objectives)
# if we don't have any data yet, use prior GP to make predictions
if self.x_dx.shape[0] == 0 or self.utility_vals is None:
pred_mean, pred_var = self._evaluate_prior(x_new)
# otherwise compute predictive mean and covariance
else:
cov_in_data = self._cov_mat_incl_deriv(x1=x_new,
x2=self.datapoints,
dx2=self.datapoints_deriv)
cov_data_in = self._cov_mat_incl_deriv(x1=self.datapoints,
dx1=self.datapoints_deriv,
x2=x_new)
cov_lamb_inv = np.linalg.inv(self.cov_mat - self.lambda_mat_inv)
pred_mean = self.prior_mean(x_new) + np.dot(
np.dot(cov_in_data, self.cov_mat_inv),
(self.utility_vals - self.prior_joint_mean(
self.datapoints, self.datapoints_deriv)))
pred_var = self._kernel(x_new, x_new) - np.dot(
np.dot(cov_in_data, cov_lamb_inv), cov_data_in)
if pointwise:
pred_var = pred_var.diagonal()
return pred_mean, pred_var
def prior_mean(self, x):
"""
Prior mean function
:param x: num_datapoints * num_objectives
:return:
"""
x = utils_data.format_data(x, self.num_objectives)
m = np.zeros(x.shape[0])
if self.prior_mean_type == 'linear':
m += np.sum(x, axis=1) / self.num_objectives
else:
TypeError('Prior mean type not understood.')
return m
def prior_mean_deriv(self, x):
"""
Derivative of prior mean function
:param x:
:return:
"""
x = utils_data.format_data(x, self.num_objectives)
if self.prior_mean_type == 'zero':
dm = np.zeros(x.shape[0])
elif self.prior_mean_type == 'linear':
dm = np.ones(x.shape[0]) / self.num_objectives
else:
TypeError('Prior mean type not understood.')
return dm
def sample(self, sample_points):
"""
Get a sample from the current GP at the given points.
:param sample_points: the points at which we want to take the sample
:return: the values of the GP sample at the input points
"""
# bring sample points in right shape
sample_points = utils_data.format_data(sample_points,
self.num_objectives)
# get the mean and the variance of the predictive (multivariate gaussian) distribution at the sample points
mean, var = self.get_predictive_params(sample_points, pointwise=False)
# sample from the multivariate gaussian with the given parameters
f_sample = self.random_state.multivariate_normal(mean, var, 1)[0]
return f_sample
def update(self, dataset, dx):
""" pairwise dataset and virtual derivative points """
self.datapoints = dataset.datapoints
self.comparisons = dataset.comparisons
# x-values at which we make virtual derivative observations
self.datapoints_deriv = utils_data.format_data(dx, self.num_objectives)
# x-values for real observations and derivative observations
self.x_dx = np.concatenate((self.datapoints, self.datapoints_deriv))
# update the covariance matrix
self.cov_mat = self._cov_mat_incl_deriv(self.datapoints,
self.datapoints_deriv)
self.cov_mat_inv = np.linalg.inv(self.cov_mat)
# update the map estimate of f
self.utility_vals = self._compute_posterior()
def _add_single_datapoint(self, new_datapoint):
"""
Add single datapoint to the existing dataset if it doesn't exist yet and return index
:param new_datapoint: new datapoint to add
:return x_new_idx: the index of the new datapoint in the existing dataset
"""
new_datapoint = utl_data.format_data(new_datapoint,
self.num_objectives)
# if the datapoint is not in our dataset yet, add it
if not utl_data.array_in_matrix(new_datapoint, self.datapoints):
self.datapoints = np.vstack((self.datapoints, new_datapoint))
new_datapoint_index = self.datapoints.shape[0] - 1
# if the datapoint is already in our dataset, find its index
else:
new_datapoint_index = self.get_index(new_datapoint)
return new_datapoint_index
def get_preference(self, x, add_noise):
# bring data into right format first
x = utils_data.format_data(x, self.num_objectives)
# EVALUATE user preferences
if self.num_objectives == 1:
y = np.polyval(self.poly, x)
else:
y = self.utility_func(x)
y = y.flatten()
assert len(y) == x.shape[0]
y = (y - self.min_y) / (self.max_y - self.min_y)
if add_noise:
# the noise is different every time the scalarisation function is called,
# but we always draw from the same distribution
noise = self.random_state.normal(0, self.std_noise, x.shape[0])
y += noise
return y
def _kernel(self, x1, x2):
x1 = utils_data.format_data(x1, self.num_objectives)
x2 = utils_data.format_data(x2, self.num_objectives)
k = 0.8**2 * np.exp(-(1. / (2. * (self.kernel_width**2))) *
np.linalg.norm(x1 - x2, axis=1)**2)
return k
