def test_basis_and_tensors():
low = np.array([0., -.5])
high = np.array([1., .5])
basis_rbf = GaussianRBF.generate([3, 3], low, high)
tensor_rbf = GaussianRBFTensor.generate([3, 3], low, high)
features_1 = Features(tensor_list=tensor_rbf)
features_2 = Features(basis_list=basis_rbf)
x = np.random.rand(10, 2) + [0., -.5]
y_1 = features_1(x)
y_2 = features_2(x)
assert np.allclose(y_1, y_2)
def experiment():
np.random.seed()
# MDP
mdp = CartPole()
# Policy
epsilon = Parameter(value=1.)
pi = EpsGreedy(epsilon=epsilon)
# Agent
basis = [PolynomialBasis()]
s1 = np.array([-np.pi, 0, np.pi]) * .25
s2 = np.array([-1, 0, 1])
for i in s1:
for j in s2:
basis.append(GaussianRBF(np.array([i, j]), np.array([1.])))
features = Features(basis_list=basis)
fit_params = dict()
approximator_params = dict(input_shape=(features.size, ),
output_shape=(mdp.info.action_space.n, ),
n_actions=mdp.info.action_space.n)
agent = LSPI(mdp.info,
pi,
approximator_params=approximator_params,
fit_params=fit_params,
features=features)
# Algorithm
core = Core(agent, mdp)
core.evaluate(n_episodes=3, render=True)
# Train
core.learn(n_episodes=100, n_episodes_per_fit=100)
# Test
test_epsilon = Parameter(0.)
agent.policy.set_epsilon(test_epsilon)
dataset = core.evaluate(n_episodes=1, quiet=True)
core.evaluate(n_steps=100, render=True)
return np.mean(episodes_length(dataset))
def test_basis():
low = np.array([0., -.5])
high = np.array([1., .5])
rbf = GaussianRBF.generate([3, 3], high, low)
features = Features(basis_list=rbf)
x = np.random.rand(10, 2) + [0., -.5]
y = features(x)
for i, x_i in enumerate(x):
assert np.all(features(x_i) == y[i])
x_1 = x[:, 0].reshape(-1, 1)
x_2 = x[:, 1].reshape(-1, 1)
assert np.all(features(x_1, x_2) == y)
for i, x_i in enumerate(zip(x_1, x_2)):
assert np.all(features(x_i[0], x_i[1]) == y[i])