def mean_var(self, x): phi_new = utility.compute_RBF(self.basis, x, self.length_scale) print phi_new intermediate = phi_new.T*self.A_inv mean = intermediate*self.phi_mat*self.Y var = self.noise*intermediate*phi_new return mean, var
def add_data(self, x, y):
"""
x is R ^ {1 * k}
y is in R
"""
if self.n == 0:
self.basis = array([x])
self.points = array([x])
self.phi_mat = mat( utility.compute_RBF(self.basis, x, self.length_scale)) # NOTE:
self.A = mat(self.noise + self.phi_mat**2)
self.A_inv = 1/self.A
self.Y = mat([y])
self.k += 1
else:
phi_new = utility.compute_RBF(self.basis, x, self.length_scale)
vec = self.A_inv*phi_new
var = phi_new.T*vec
self.A_inv = self.A_inv - vec*vec.T/(1+var)
if self.noise*var <= self.vu:
"""
Not adding basis.
"""
print var
self.A = self.A + phi_new*phi_new.T
self.phi_mat = hstack([self.phi_mat, phi_new])
else:
"""
Add basis.
"""
phi_new_2 = utility.compute_RBF(self.points, x, self.length_scale)
self.k += 1
d = utility.compute_RBF([x], x, self.length_scale)
left_vec = self.phi_mat*phi_new_2 + float(d)*phi_new
last = phi_new_2.T*phi_new_2 + d**2 + self.noise
vec = self.A_inv*left_vec
var = left_vec.T*vec
common_factor = 1.0/(last - var)
# Update A
self.A = vstack([hstack([self.A + phi_new*phi_new.T, left_vec ]), hstack([left_vec.T, last])])
# Update A_inv
self.A_inv = vstack([hstack([self.A_inv + (vec*vec.T)*float(common_factor), -float(common_factor)*vec]), hstack([-float(common_factor)*vec.T, common_factor])])
# Update self.phi_mat
self.phi_mat = vstack([hstack([self.phi_mat, phi_new]), hstack([phi_new_2.T, d])])
# Add to basis
self.basis = r_[self.basis, [x]]
#print linalg.norm(linalg.inv(self.phi_mat*self.phi_mat.T+self.noise*eye(self.n+1)) - linalg.inv(self.A))
#print linalg.norm(linalg.inv(self.A) - self.A_inv)
# Add to points
self.points = r_[self.points, [x]]
# Add to self.Y
self.Y = vstack([self.Y, mat([y])])
self.n += 1
