x0 = [np.random.uniform(-0.005, 0.005), np.pi, 0, 0]
return x0
x0 = [0, np.pi, 0, 0]
t = 5 # time of an episode
dt = args.time_step # time step-size
env = CartPole(c_k, x0, dt)
path = args.path + str(args.exp_id) + '/'
rl_algorithm = MBRL(env, t, dt,
test_t=args.test_t,
path=path,
horizon=2.,
fcost=None,
warm_up_episodes=args.warm_up_episodes,
use_mpc=args.use_mpc,
ilqr_print=True,
ilqr_save=False,
aggregation_interval=args.agg,
training_epochs=args.epochs,
weight_decay=args.weight_decay,
prediction_error_bound=args.pred_err_bound,
dyn_lr=1e-5)
#rl_algorithm.load()
rl_algorithm.run_learning(50)
# define the function, that represents the initial value distribution p(x_0)
def p_x0():
x0 = [np.random.uniform(-0.005, 0.005), np.pi, np.pi, 0, 0, 0]
return x0
t = 3.5 # time of an episode
dt = args.time_step # time step-size
env = CartPoleDouble(c_k, p_x0, dt)
path = args.path + str(args.exp_id)+'/'
rl_algorithm = MBRL(env, t, dt,
path=path,
horizon=2.,
fcost=c_N,
warm_up_episodes=args.warm_up_episodes,
use_mpc=args.use_mpc,
ilqr_print=True,
ilqr_save=True,
aggregation_interval=args.agg,
training_epochs=args.epochs,
weight_decay=args.weight_decay,
data_noise=args.data_noise)
if args.data_set != '':
rl_algorithm.D_rand.load(args.data_set)
rl_algorithm.run_learning(args.episodes)
t = args.t # time of an episode
dt = args.time_step # time step-size
env = Pendulum(c_k, x0, dt)
env.uMax = env.uMax / 3.5 * 5
path = args.path + str(args.exp_id) + '/'
rl_algorithm = MBRL(env,
t,
dt,
path=path,
horizon=t,
fcost=c_N,
test_t=args.test_t,
warm_up_episodes=args.warm_up_episodes,
use_mpc=args.use_mpc,
ilqr_print=True,
ilqr_save=True,
aggregation_interval=args.agg,
training_epochs=args.epochs,
weight_decay=args.weight_decay,
data_noise=args.data_noise,
prediction_error_bound=args.pred_err_bound,
maxIters=100,
sparse_dyn=True)
rl_algorithm.run_learning(args.episodes)