def cmc_config():
# run parameters
nb_bootstrap = 10
nb_explorations = 50
nb_tests = 100
nb_timesteps = 1000
offline_eval = (1e6, 10) # (x,y): y evaluation episodes every x (done offline)
# controller parameters
hidden_sizes = []
controller_tmp = 1.
activation = 'relu'
subset_obs = range(2)
norm_values = None
scale = np.array([[-1.2,0.6],[-0.07,0.07]])
controller = NNController(hidden_sizes, controller_tmp, subset_obs, 1, norm_values, scale, activation)
nb_weights = controller.nb_weights
# representer
representer = CMCRepresenter()
initial_space = representer.initial_space
goal_space = representer.initial_space # space in which goal are sampled
nb_rep = representer.dim
engineer_goal = np.array([0.5, 1.5, 5.]) # engineer goal
# scale engineer goal to [-1,1]^N
engineer_goal = scale_vec(engineer_goal, initial_space)
# inverse model
knn = KNNRegressor(n_neighbors=1)
# exploration_noise
noise = 0.1
return nb_bootstrap, nb_explorations, nb_tests, nb_timesteps, offline_eval, \
controller, representer, nb_rep, engineer_goal, goal_space, initial_space, knn, noise, nb_weights
def cheetah_config():
# run parameters
nb_bootstrap = 200
nb_explorations = 1000
nb_tests = 100
nb_timesteps = 1000
offline_eval = (1e6, 10
) # (x,y): y evaluation episodes every x (done offline)
# controller parameters
hidden_sizes = []
controller_tmp = 0.15
activation = 'relu'
subset_obs = [2, 3, 4, 5, 6, 7, 11, 12, 13, 14, 15, 16]
norm_values = np.array([[
-0.18, 0.064, 0.14, 0.17, 0.079, -0.082, -0.08, -0.084, 0.30, -0.011,
0.0078, 0.0065, -0.055, 0.014, -0.011, 0.073, 0.054
],
[
0.073, 0.25, 0.22, 0.22, 0.25, 0.35, 0.27,
0.25, 0.64, 0.38, 1.1, 2.1, 2.9, 3.5, 3.9, 3.5,
3.4
]]) # to zscore observations
scale = None
controller = NNController(hidden_sizes, controller_tmp, subset_obs, 6,
norm_values, scale, activation)
nb_weights = controller.nb_weights
# representer
representer = CheetahRepresenter()
initial_space = representer.initial_space
goal_space = np.array([[2., 6.], [-0.5,
0.]]) # space in which goal are sampled
engineer_goal = np.array([6., -0.2]) # engineer goal
# scale engineer goal to [-1,1]^N
from gep_utils import scale_vec
engineer_goal = scale_vec(engineer_goal, initial_space)
nb_rep = representer.dim
# inverse model
knn = KNNRegressor(n_neighbors=1)
# exploration_noise
noise = 0.1
return nb_bootstrap, nb_explorations, nb_tests, nb_timesteps, offline_eval, \
controller, representer, nb_rep, engineer_goal, goal_space, initial_space, knn, noise, nb_weights
def cmc_config():
# run parameters
nb_bootstrap = 10
nb_tests = 100
nb_timesteps = 1000
offline_eval = (1e6, 10
) # (x,y): y evaluation episodes every x (done offline)
# controller parameters
hidden_sizes = []
dim_state = 2
dim_act = 1
controller_tmp = 1.
activation = 'relu'
scale = np.array([[-1.2, 0.6], [-0.07, 0.07]])
controller = NNController(dim_state, hidden_sizes, dim_act, controller_tmp,
scale, activation)
nb_weights = controller.nb_weights
# representer
representer = CMCRepresenter()
initial_space = representer.initial_space
goal_space = representer.initial_space # space in which goal are sampled
nb_rep = representer.dim
engineer_goal = np.array([0.5, 1.5, 5.]) # engineer goal
# scale engineer goal to [-1,1]^N
engineer_goal = scale_vec(engineer_goal, initial_space)
# inverse model
knn = KNNRegressor(n_neighbors=1)
# exploration_noise
noise = 0.1
# policy sampling method
# 'uniform': in [-1, 1]
# 'normal' (0, 1)
policy_sampling_method = 'uniform'
# best policy selection
# to find best policy in archive, use:
# 'goal': the closest to some engineered goal in the behavioral space,
# 'perf': the policy that obtained the highest performance
best_policy_selection = 'goal'
return nb_bootstrap, nb_tests, nb_timesteps, offline_eval, controller, representer, \
policy_sampling_method, nb_rep, engineer_goal, goal_space, initial_space, knn, noise, nb_weights, best_policy_selection
def kobuki_config():
# run parameters
nb_bootstrap = 200
nb_explorations = 1000
nb_tests = 10
#nb_timesteps = random.randint(1,50)
#print('Timesteps: ' + str(nb_timesteps))
nb_timesteps = 30
offline_eval = (1e6, 10)
# controller parameters
hidden_sizes = []
controller_tmp = 1.
activation = 'relu'
subset_obs = range(6)
norm_values = None
#scale = np.vstack([np.array([[-1.0,1.0],]*2)])
scale = np.array([[-1.0, 1.0], [-1.0, 1.0], [-1.0, 1.0], [-1.0, 1.0],
[0.0, 1.0], [0.0, 1.0]])
controller = NNController(hidden_sizes, controller_tmp, subset_obs, 2,
norm_values, scale, activation)
nb_weights = controller.nb_weights
# representer
representer = KobukiRepresenter()
initial_space = representer.initial_space
goal_space = representer.initial_space
nb_rep = representer.dim
engineer_goal = np.random.uniform(-0.2, 0.2, (2, ))
# scale engineer goal to[-1, 1]^N
engineer_goal = scale_vec(engineer_goal, initial_space)
# inverse model
knn = KNNRegressor(n_neighbors=1)
# exploration_noise
noise = 0.1
return nb_bootstrap, nb_explorations, nb_tests, nb_timesteps, offline_eval, \
controller, representer, nb_rep, engineer_goal, goal_space, initial_space, knn, noise, nb_weights