def evaluate_MSE(self, true_func, res=30):
data = generate_grid(-2.0, 2.0, res)
self.model = GP(self.x_train[2], self.y_train[2], self.kernel)
m_pred, s_pred = self.model.predict(data)
m_true = np.apply_along_axis(true_func, axis=1,
arr=data).reshape(-1, 1)
return np.sum((m_pred - m_true)**2) / (res * res)
def compute_entropy(self, res=21):
#print(self.x_train[2], self.y_train[2])
self.model = GP(self.x_train[2], self.y_train[2], self.kernel)
x_candidates = generate_grid(-2.0, 2.0, res)
mean, var = self.model.predict(x_candidates)
self.entropy = np.sum(np.log(var))
#self.entropy = mean + 0.3 * var
#print(self.entropy)
return self.entropy
def compute_belief_variance(self):
priors = [[] for i in range(self.num_features)]
for i in range(self.num_features):
priors[i] = GP(self.x_train[i], self.y_train[i], self.kernel)
#plotGP(priors[i], x_train[i], 'Feature '+str(i))
m_obs = [[] for i in range(self.num_features)]
m_exp = [[] for i in range(self.num_features)]
s_obs = [[] for i in range(self.num_features)]
### sample variable distributions
#for x in range(len(x_candidates)):
for i in range(self.num_features):
pred_obs = priors[2].predict(self.x_train[i])
m_obs[i], s_obs[i] = pred_obs[0].flatten(), pred_obs[1].flatten()
m_exp[i] = self.y_train[i]
#print(pred[0].flatten())
self.count = np.ones(self.num_features)
for i in range(self.num_features):
for j in range(len(self.x_train[i])):
if i != 2 and s_obs[i][
j] < 0.8: # if sufficiently confident about feature of interest where the side feature has been observed
self.count[
i] += 1 # we know more about the relationship between those features
self.feature_var = np.array([1.0 / math.sqrt(n) for n in self.count])
def infer_joint_distribution(self, res):
### train GP priors
priors = [[] for i in range(self.num_features)]
for i in range(self.num_features):
priors[i] = GP(self.x_train[i], self.y_train[i], self.kernel)
#plotGP(priors[i], x_train[i], 'Feature '+str(i))
m = np.ndarray((self.num_features, res * res))
s = np.ndarray((self.num_features, res * res))
x_candidates = generate_grid(-2, 2, res)
### sample variable distributions
for x in range(len(x_candidates)):
for i in range(self.num_features):
pred = priors[i].predict(np.array([x_candidates[x]]))
m[i, x], s[i, x] = pred[0][0][0], pred[1][0][0]
### compute weighted covariance
W = self.generate_weights(s)
X = np.array([m[i] - np.mean(m[i]) for i in range(self.num_features)])
w_cov = np.cov(X, aweights=W)
### use lasso to estimate a sparse precision matrix? (GGM model estimation)
### construct correlated joint distribution GP
joint_distribution = MOGP(self.x_train, self.y_train, w_cov,
self.kernel)
#plotMOGP(joint_distribution, x_train, 2, 'Output 2')
return joint_distribution
def infer_joint_distribution(self, res):
### train GP priors
priors = [[] for i in range(self.num_features)]
for i in range(self.num_features):
priors[i] = GP(self.x_train[i], self.y_train[i], self.kernel)
#plotGP(priors[i], x_train[i], 'Feature '+str(i))
m_obs = [[] for i in range(self.num_features)]
m_exp = [[] for i in range(self.num_features)]
s_obs = [[] for i in range(self.num_features)]
x_candidates = generate_grid(-2, 2, res)
### sample variable distributions
#for x in range(len(x_candidates)):
for i in range(self.num_features):
pred_obs = priors[2].predict(self.x_train[i])
m_obs[i], s_obs[i] = pred_obs[0].flatten(), pred_obs[1].flatten()
m_expert[i] = self.y_train[i]
#print(pred[0].flatten())
self.count = np.zeros(self.num_features)
m_obs_trust = [[] for i in range(self.num_features)]
m_exp_trust = [[] for i in range(self.num_features)]
for i in range(self.num_features):
for j in range(len(self.x_train[i])):
if s_obs[i][
j] < 0.3: # if sufficiently confident about feature of interest where the side feature has been observed
count[
i] += 1 # we know more about the relationship between those features
self.feature_var = np.array([1.0 / math.sqrt(n) for n in count])
#print(m[0, :])
#for x in range(len(x_candidates)):
# for i in range(self.num_features):
# pred = priors[i].predict(np.array([x_candidates[x]]))
# m[i, x], s[i, x] = pred[0][0][0], pred[1][0][0]
for i in range(self.num_features):
pred = priors[i].predict(x_candidates)
m[i], s[i] = pred[0].flatten(), pred[1].flatten()
### compute weighted covariance
W = self.generate_weights(s)
X = np.array([m[i] - np.mean(m[i]) for i in range(self.num_features)])
w_cov = np.cov(X, aweights=W)
### use lasso to estimate a sparse precision matrix? (GGM model estimation)
### construct correlated joint distribution GP
joint_distribution = MOGP(self.x_train, self.y_train, w_cov,
self.kernel)
#plotMOGP(joint_distribution, x_train, 2, 'Output 2')
return joint_distribution
def infer_joint_distribution(self, res):
### train GP priors
priors = [[] for i in range(self.num_features)]
for i in range(self.num_features):
priors[i] = GP(self.x_train[i], self.y_train[i], self.kernel)
#plotGP(priors[i], x_train[i], 'Feature '+str(i))
### construct correlated joint distribution GP
joint_distribution = MOGP(self.x_train, self.y_train, self.cov, self.kernel)
#plotMOGP(joint_distribution, x_train, 2, 'Output 2')
return joint_distribution
def infer_independent_distribution(self, feature, res):
independent_distribution = GP(self.x_train[feature],
self.y_train[feature], self.kernel)
return independent_distribution
class GaussianProcessBeliefModel(): # InferenceModel
def __init__(self):
self.res = 20
self.kernel = generate_rbfkern(2, 1.0, 0.3)
self.entropy = 0
def copy(self):
newMe = GaussianProcessBeliefModel()
newMe.x_train, newMe.y_train, newMe.num_features, newMe.entropy = self.x_train.copy(
), self.y_train.copy(), self.num_features, self.entropy
return newMe
def load_environment(self, env, start_loc=[0, 0]):
self.x_train, self.y_train, self.num_features = env.load_prior_data(
start_loc)
#self.entropy = -1 * len(self.x_train)
#print(self.entropy)
def update(self, x, y, feature):
#print(np.min(abs(np.sum(self.x_train[feature] - [x[0:2]], axis=1))))
#print('a')
#print([x[0:2]])
#print(self.x_train[feature])
#print(self.x_train[feature] - [x[0:2]])
#print(np.sum(abs(self.x_train[feature] - [x[0:2]]), axis=1))
self.x_train[feature] = np.append(self.x_train[feature], [x[0:2]],
axis=0)
self.y_train[feature] = np.append(self.y_train[feature], y)
def update_static(self, x, y, feature):
self.x_train[feature] = np.append(self.x_train[feature], [x[0:2]],
axis=0)
self.y_train[feature] = np.append(self.y_train[feature], y)
def observe(self, x):
feature = int(x[0][2])
#if x[0][0:2] in self.x_train[feature]:
# print(np.where(self.x_train[feature] == x[0][0:2]))
# return self.y_train[np.where(self.x_train[feature] == x[0][0:2])]
#else:
return 0
def infer_independent_distribution(self, feature, res):
independent_distribution = GP(self.x_train[feature],
self.y_train[feature], self.kernel)
return independent_distribution
def display(self, feature, title):
independent_distribution = self.infer_independent_distribution(
feature=feature, res=20)
plotGP(independent_distribution,
self.x_train[feature],
title=title,
res=20)
def generate_weights(self, S):
W = np.reciprocal(np.sqrt(S[2, :]))
return W
def compute_entropy(self, res=21):
#print(self.x_train[2], self.y_train[2])
self.model = GP(self.x_train[2], self.y_train[2], self.kernel)
x_candidates = generate_grid(-2.0, 2.0, res)
mean, var = self.model.predict(x_candidates)
self.entropy = np.sum(np.log(var))
#self.entropy = mean + 0.3 * var
#print(self.entropy)
return self.entropy
def evaluate_MSE(self, true_func, res=30):
data = generate_grid(-2.0, 2.0, res)
self.model = GP(self.x_train[2], self.y_train[2], self.kernel)
m_pred, s_pred = self.model.predict(data)
m_true = np.apply_along_axis(true_func, axis=1,
arr=data).reshape(-1, 1)
return np.sum((m_pred - m_true)**2) / (res * res)