64 lines
2.3 KiB
Python
64 lines
2.3 KiB
Python
import numpy as np
|
|
from scipy.stats import norm
|
|
|
|
|
|
class PreferenceExpectedImprovement:
|
|
def __init__(self, nr_samples, nr_dims, lower_bound, upper_bound, seed=None):
|
|
|
|
self.nr_samples = nr_samples
|
|
self.nr_dims = nr_dims
|
|
# check if upper_bound and lower_bound are numpy arrays of shape (nr_dims, 1) or (nr_dims,) or if they are floats
|
|
|
|
self.upper_bound = upper_bound
|
|
self.lower_bound = lower_bound
|
|
|
|
self.initial_variance = 5.0
|
|
|
|
self.user_model = None
|
|
self.proposal_model_mean = np.array((nr_dims, 1))
|
|
self.proposal_model_covariance = np.diag(np.ones((nr_dims, )) * self.initial_variance)
|
|
|
|
self.rng = np.random.default_rng(seed=seed)
|
|
|
|
def initialize(self):
|
|
pass
|
|
|
|
def rejection_sampling(self):
|
|
samples = np.empty((0, self.nr_dims))
|
|
while samples.shape[0] < self.nr_samples:
|
|
# sample from the multi variate gaussian distribution
|
|
sample = self.rng.multivariate_normal(
|
|
self.proposal_model_mean,
|
|
self.proposal_model_covariance
|
|
)
|
|
|
|
# check if the sample is within the bounds
|
|
if np.all(sample >= self.lower_bound) and np.all(sample <= self.upper_bound):
|
|
samples = np.append(samples, [sample], axis=0)
|
|
|
|
return samples
|
|
|
|
|
|
def expected_improvement(self):
|
|
pass
|
|
|
|
def update_user_preference_model(self, preferred_input, preference_array):
|
|
# Update mean to reflect preferred input
|
|
self.user_model_mean = preferred_input
|
|
|
|
initial_variance = np.ones((self.nr_dims, )) * self.initial_variance
|
|
reduced_variance = initial_variance / 10.0
|
|
variances = np.where(preference_array, reduced_variance, initial_variance)
|
|
self.user_model_covariance = np.diag(variances)
|
|
|
|
def update_proposal_model(self, alpha=0.5):
|
|
# Update proposal model to be a weighted average of the current proposal model and the user model
|
|
self.proposal_model_mean = alpha * self.proposal_model_mean + (1 - alpha) * self.user_model_mean
|
|
self.proposal_model_covariance = alpha * self.proposal_model_covariance + (1 - alpha) * self.user_model_covariance
|
|
|
|
|
|
if __name__ == '__main__':
|
|
acquisition = PreferenceExpectedImprovement(10, 2, -1.0, 1.0)
|
|
|
|
|