def approx_direction(self, x):
"""
Calculate - H^{-1} * g
"""
g = self.grad(x)
H = self.hess(x)
try:
L = np.linalg.cholesky(H)
except np.linalg.LinAlgError:
raise ValueError("The hessian is not positive definite.")
# Solve L * y = b
y = ut.forward_sub(L, -g)
# Solve U * s = y
d = ut.backward_sub(L.T, y)
return d
def approx_direction(self, x):
g = self.grad(x)
try:
gamma = g - self.g_old
except:
gamma = np.ones(self.dim)
self.g_old = g.copy()
H = self.hess
P, L, U = lu(H)
# d = np.linalg.solve(H, -g)
# Solve L * y = b
y = ut.forward_sub(L, -g)
# Solve U * s = y
d = ut.backward_sub(U, y)
d = np.dot(np.linalg.inv(H), -g)
return d, gamma
