def test_state_hessian(self):
"""
"""
g_xs = np.ones((1, 3)) * 0.5
xs = np.ones((1, 3))
cost_hess =\
TwoWheeledConfigModule.hessian_cost_fn_state(xs, g_xs)
# numeric grad
eps = 1e-4
expected_hess = np.zeros((1, 3, 3))
for i in range(3):
tmp_x = xs.copy()
tmp_x[0, i] = xs[0, i] + eps
forward = \
TwoWheeledConfigModule.gradient_cost_fn_state(
tmp_x, g_xs, terminal=False)
tmp_x = xs.copy()
tmp_x[0, i] = xs[0, i] - eps
backward = \
TwoWheeledConfigModule.gradient_cost_fn_state(
tmp_x, g_xs, terminal=False)
expected_hess[0, :, i] = (forward - backward) / (2. * eps)
assert cost_hess == pytest.approx(expected_hess)
def test_state_gradient(self):
"""
"""
xs = np.ones((1, 3))
g_xs = np.ones((1, 3)) * 0.5
cost_grad =\
TwoWheeledConfigModule.gradient_cost_fn_state(xs, g_xs)
# numeric grad
eps = 1e-4
expected_grad = np.zeros((1, 3))
for i in range(3):
tmp_x = xs.copy()
tmp_x[0, i] = xs[0, i] + eps
forward = \
TwoWheeledConfigModule.state_cost_fn(tmp_x, g_xs)
forward = np.sum(forward)
tmp_x = xs.copy()
tmp_x[0, i] = xs[0, i] - eps
backward = \
TwoWheeledConfigModule.state_cost_fn(tmp_x, g_xs)
backward = np.sum(backward)
expected_grad[0, i] = (forward - backward) / (2. * eps)
assert cost_grad == pytest.approx(expected_grad)