# @Filename: mf_pendulum.py # @Date: 2019-06-16-18-38 # @Author: Hany Abdulsamad # @Contact: [email protected] import gym from trajopt.gps import MFGPS # pendulum env env = gym.make('Pendulum-TO-v0') env._max_episode_steps = 150 alg = MFGPS(env, nb_steps=150, kl_bound=25., init_ctl_sigma=1.0, activation=range(100, 150)) # run gps trace = alg.run(nb_episodes=10) # plot dists alg.plot() # execute and plot nb_episodes = 25 data = alg.sample(nb_episodes, stoch=False) import matplotlib.pyplot as plt plt.figure()
import gym
from trajopt.gps import MFGPS
# lqr task
env = gym.make('LQR-TO-v0')
env._max_episode_steps = 100
alg = MFGPS(env, nb_steps=100,
kl_bound=10.,
init_ctl_sigma=50.)
# run gps
trace = alg.run(nb_episodes=10, nb_iter=5)
# plot dists
alg.plot()
# execute and plot
nb_episodes = 25
data = alg.sample(nb_episodes, stoch=False)
import matplotlib.pyplot as plt
plt.figure()
for k in range(alg.dm_state):
plt.subplot(alg.dm_state + alg.dm_act, 1, k + 1)
plt.plot(data['x'][k, ...])
for k in range(alg.dm_act):
plt.subplot(alg.dm_state + alg.dm_act, 1, alg.dm_state + k + 1)
# pendulum task
env = gym.make('DoublePendulum-TO-v0')
env._max_episode_steps = 100
env.unwrapped.dt = 0.05
env.seed(1337)
# prior = {'K': 1e-3, 'psi': 1e-8, 'nu': 0.1}
prior = {'K': 1e-6}
solver = MFGPS(env,
nb_steps=100,
init_state=env.init(),
init_action_sigma=10.,
kl_bound=1e1,
slew_rate=False,
action_penalty=1,
prior=prior,
activation={
'mult': 1.,
'shift': 80
})
# run gps
trace = solver.run(nb_learning_episodes=25,
nb_evaluation_episodes=25,
nb_iter=50,
verbose=True)
# execute and plot
data = solver.rollout(25, stoch=True)
import warnings
warnings.filterwarnings("ignore")
np.random.seed(1337)
# lqr task
env = gym.make('LQR-TO-v1')
env._max_episode_steps = 100000
env.seed(1337)
prior = {'K': 1e-6, 'psi': 1e6, 'nu': 0.1}
alg = MFGPS(env, nb_steps=60,
init_state=env.init(),
init_action_sigma=100.,
kl_bound=1e0,
prior=prior)
# run gps
trace = alg.run(nb_learning_episodes=25,
nb_evaluation_episodes=25,
nb_iter=25, verbose=True)
# plot dists
alg.plot_distributions()
# execute and plot
data = alg.rollout(25, stoch=True)
plt.figure()
import gym
from trajopt.gps import MFGPS
# pendulum env
env = gym.make('Pendulum-TO-v0')
env._max_episode_steps = 100
env.unwrapped._dt = 0.05
alg = MFGPS(env,
nb_steps=100,
kl_bound=1.,
init_ctl_sigma=2.5,
activation={
'shift': 90,
'mult': 2.
})
# run gps
trace = alg.run(nb_episodes=10)
# plot dists
alg.plot()
# execute and plot
nb_episodes = 25
data = alg.sample(nb_episodes, stoch=False)
import matplotlib.pyplot as plt
plt.figure()
for k in range(alg.dm_state):
prior = {'K': 1e-6}
# mass-damper
rgps = MFRGPS(env, nb_steps=100,
policy_kl_bound=5.,
param_kl_bound=1e1,
init_state=env.init(),
init_action_sigma=100.,
prior=prior)
np.random.seed(1337)
env.seed(1337)
rgps.run(nb_iter=15, nb_learning_episodes=10,
nb_evaluation_episodes=100, verbose=True)
gps = MFGPS(env, nb_steps=100,
kl_bound=5.,
init_state=env.init(),
init_action_sigma=100.,
dyn_prior=prior)
np.random.seed(1337)
env.seed(1337)
gps.run(nb_iter=15, nb_learning_episodes=10,
nb_evaluation_episodes=100, verbose=True)
gps.plot()
init_action_sigma=25.,
action_penalty=1e-1,
activation={'mult': 1., 'shift': 80},
prior=prior)
rgps_trace = rgps.run(nb_learning_episodes=10,
nb_evaluation_episodes=100,
nb_iter=50, verbose=True)
np.random.seed(1337)
env.seed(1337)
gps = MFGPS(env, nb_steps=100,
kl_bound=1e1,
init_state=env.init(),
init_action_sigma=25.,
action_penalty=1e-1,
activation={'mult': 1., 'shift': 80},
dyn_prior=prior)
# run gps
gps_trace = gps.run(nb_learning_episodes=10,
nb_evaluation_episodes=100,
nb_iter=50, verbose=True)
np.random.seed(1337)
env.seed(1337)
rgps_data = rgps.rollout(100, stoch=False)
print("Cost of Rbst. Ctl.:", np.mean(np.sum(rgps_data['c'], axis=0)))
np.random.seed(1337)