def run_irl(world, car, reward, theta, data):
def gen():
for point in data:
for c, x0, u in zip(world.cars, point['x0'], point['u']):
c.traj.x0.set_value(x0)
for cu, uu in zip(c.traj.u, u):
cu.set_value(uu)
yield
r = car.traj.reward(reward)
g = utils.grad(r, car.traj.u)
H = utils.hessian(r, car.traj.u)
I = tt.eye(utils.shape(H)[0])
reg = utils.vector(1)
reg.set_value([1e-1])
H = H - reg[0] * I
L = tt.dot(g, tt.dot(tn.MatrixInverse()(H), g)) + tt.log(tn.Det()(-H))
for _ in gen():
pass
optimizer = utils.Maximizer(L, [theta],
gen=gen,
method='gd',
eps=0.1,
debug=True,
iters=1000,
inf_ignore=10)
optimizer.maximize()
print theta.get_value()
def minus_two_log_gauss_likelihood_2D(residuals, covariance_values):
"""computes the -2 log gaussian likelihood (ignoring the constant term)
from the given residuals and covariance values in the form of a
1D array: [variance1, covariance, variance2]"""
cov = covariance_matrix_2D(covariance_values)
det = L.Det()(cov)
precis = L.MatrixInverse()(cov)
term1 = T.dot(T.transpose(residuals), T.dot(precis, residuals))
term2 = T.log(det)
return term1 + term2