def test_penalize_velocity():
env = OptimumTrajectory(x0=np.zeros(1),
g=np.zeros(1),
dt=1.0,
penalty_vel=6.0)
env.init()
n_outputs = env.get_num_outputs()
n_inputs = env.get_num_inputs()
xva = np.empty(n_outputs)
env.reset()
assert_false(env.is_evaluation_done())
env.get_outputs(xva)
xva = np.array([0.0, 1.0, 0.0])
env.set_inputs(xva)
env.step_action()
assert_false(env.is_evaluation_done())
env.get_outputs(xva)
xva = np.array([0.0, 1.0, 0.0])
env.set_inputs(xva)
env.step_action()
assert_true(env.is_evaluation_done())
rewards = env.get_feedback()
assert_array_almost_equal(rewards, -6.0 * np.ones(2))
def test_dimensions():
env = OptimumTrajectory(x0=np.zeros(3), g=np.ones(3))
env.init()
n_outputs = env.get_num_outputs()
n_inputs = env.get_num_inputs()
assert_equal(n_outputs, n_inputs)
def test_penalize_start():
env = OptimumTrajectory(x0=np.zeros(1),
g=np.ones(1),
dt=1.0,
penalty_start_dist=4.0)
env.init()
n_outputs = env.get_num_outputs()
n_inputs = env.get_num_inputs()
xva = np.empty(n_outputs)
env.reset()
assert_false(env.is_evaluation_done())
env.get_outputs(xva)
xva[:] = 1.0
env.set_inputs(xva)
env.step_action()
assert_false(env.is_evaluation_done())
env.get_outputs(xva)
assert_array_equal(xva, np.ones(3))
env.set_inputs(xva)
env.step_action()
assert_true(env.is_evaluation_done())
rewards = env.get_feedback()
assert_array_almost_equal(rewards, np.array([-4.0, 0.0]))
def test_penalize_obstacles():
obstacles = np.array([[0.0, 0.0]])
env = OptimumTrajectory(dt=1.0, penalty_obstacle=8.0, obstacles=obstacles)
env.init()
n_outputs = env.get_num_outputs()
n_inputs = env.get_num_inputs()
xva = np.empty(n_outputs)
env.reset()
assert_false(env.is_evaluation_done())
env.get_outputs(xva)
env.set_inputs(xva)
env.step_action()
assert_false(env.is_evaluation_done())
env.get_outputs(xva)
env.set_inputs(xva)
env.step_action()
assert_true(env.is_evaluation_done())
rewards = env.get_feedback()
assert_array_almost_equal(rewards, -8.0 * np.ones(2))
dt,
n_features,
learn_covariance=True,
use_covar=True)
# init linear to have a guess
beh.init(4, 4)
beh.set_meta_parameters(["g", "x0"], [g, x0])
beh.imitate(np.tile(np.linspace(0, 1, 101), 2).reshape((2, 101, -1)))
env = OptimumTrajectory(x0,
g,
execution_time,
dt,
obstacles,
penalty_goal_dist=10000.0,
penalty_start_dist=10000.0,
penalty_obstacle=1000.0,
penalty_length=10.,
hide_acc_from_interface=True,
use_covar=True)
opt = CMAESOptimizer(variance=0.1**2,
random_state=0,
initial_params=beh.get_params())
bs = BlackBoxSearch(beh, opt)
controller = Controller(environment=env,
behavior_search=bs,
n_episodes=n_episodes,
record_inputs=True)
rewards = controller.learn(["x0", "g"], [x0, g])
from bolero.controller import Controller
n_task_dims = 2
obstacles = [np.array([0.5, 0.5]), np.array([0.6, 0.8]), np.array([0.8, 0.6])]
x0 = np.zeros(n_task_dims)
g = np.ones(n_task_dims)
execution_time = 1.0
dt = 0.01
n_features = 10
n_episodes = 500
beh = DMPBehavior(execution_time, dt, n_features)
env = OptimumTrajectory(x0,
g,
execution_time,
dt,
obstacles,
penalty_goal_dist=1.0,
penalty_obstacle=1000.0,
penalty_acc=1.0)
opt = CMAESOptimizer(variance=100.0**2, random_state=0)
bs = BlackBoxSearch(beh, opt)
controller = Controller(environment=env,
behavior_search=bs,
n_episodes=n_episodes,
record_inputs=True)
rewards = controller.learn(["x0", "g"], [x0, g])
controller.episode_with(bs.get_best_behavior(), ["x0", "g"], [x0, g])
X = np.asarray(controller.inputs_[-1])
X_hist = np.asarray(controller.inputs_)