parallel_sampler.initialize(n_parallel=1)
parallel_sampler.set_seed(0)
policy = QMDPPolicy(env_spec=env.spec,
name="QMDP",
qmdp_param=env._wrapped_env.params)
baseline = LinearFeatureBaseline(env_spec=env.spec)
with tf.Session() as sess:
# writer = tf.summary.FileWriter(logdir=log_dir,)
algo = VPG_t(
env=env,
policy=policy,
baseline=baseline,
batch_size=2048, #2*env._wrapped_env.params['traj_limit'],
max_path_length=env._wrapped_env.params['traj_limit'],
n_itr=10000,
discount=0.95,
step_size=0.01,
record_rewards=True,
transfer=False,
)
algo.train(sess)
# tf.summary.merge_all()
# print(sess.graph)
# writer.add_graph(sess.graph)
# writer.close()
name="QMDP",
qmdp_param=env._wrapped_env.params)
baseline = LinearFeatureBaseline(env_spec=env.spec)
with tf.Session() as sess:
writer = tf.summary.FileWriter(logdir=log_dir, )
algo = VPG_t(
env=env,
policy=policy,
baseline=baseline,
batch_size=2 * env._wrapped_env.params['traj_limit'],
max_path_length=env._wrapped_env.params['traj_limit'],
n_itr=1,
discount=0.95,
step_size=0.01,
record_rewards=True,
transfer=False,
# optimizer=ConjugateGradientOptimizer(hvp_approach=FiniteDifferenceHvp(base_eps=1e-5))
# optimizer = PenaltyLbfgsOptimizer()
# Uncomment both lines (this and the plot parameter below) to enable plotting
# plot=True,
)
algo.train(sess)
# tf.summary.merge_all()
# print(sess.graph)
writer.add_graph(sess.graph)
writer.close()