sess = tf.Session()
sess.__enter__()
# Instantiate the env
data = joblib.load("../CartPole/ControlPolicy/itr_5.pkl")
sut = data['policy']
reward_function = ASTRewardS()
simulator = CartpoleSimulator(sut=sut,
max_path_length=max_path_length,
use_seed=False,
nd=1)
env = TfEnv(
ASTEnv(
open_loop=open_loop,
simulator=simulator,
fixed_init_state=True,
s_0=[0.0, 0.0, 0.0 * math.pi / 180, 0.0],
reward_function=reward_function,
))
# Training
with open(osp.join(args.log_dir, 'total_result.csv'), mode='w') as csv_file:
fieldnames = ['step_count']
for i in range(top_k):
fieldnames.append('reward ' + str(i))
writer = csv.DictWriter(csv_file, fieldnames=fieldnames)
writer.writeheader()
for trial in range(args.n_trial):
# Create the logger
log_dir = args.log_dir + '/' + str(trial)
tf.set_random_seed(seed)
with tf.Session() as sess:
# Create env
data = joblib.load("../CartPole/ControlPolicy/itr_5.pkl")
sut = data['policy']
reward_function = ASTRewardS()
simulator = CartpoleSimulator(sut=sut, max_path_length=100, use_seed=False)
env = ASTEnv(open_loop=False,
simulator=simulator,
fixed_init_state=True,
s_0=[0.0, 0.0, 0.0 * math.pi / 180, 0.0],
reward_function=reward_function,
)
env = TfEnv(env)
# Create policy
policy = DeterministicMLPPolicy(
name='ast_agent',
env_spec=env.spec,
hidden_sizes=(64, 32),
output_nonlinearity=tf.nn.tanh,
)
params = policy.get_params()
sess.run(tf.variables_initializer(params))
# Instantiate the garage objects
baseline = ZeroBaseline(env_spec=env.spec)
# optimizer = ConjugateGradientOptimizer(hvp_approach=FiniteDifferenceHvp(base_eps=1e-5))
