In [1]:

    

import tensorflow as tf


    

In [2]:

    

from drl.ddpg import DDPG
from drl.exploration import OrnSteinUhlenbeckNoise, WhiteNoise, LinearDecay
from drl.utilities import Statistics
from drl.env.arm import TwoLinkArm


    

    




Using TensorFlow backend.

In [3]:

    

ENV_NAME = "TwoLinkArm"
ALGO_NAME = "DDPG"
SAVE = False

SETTINGS = {
    'learning_rate_actor': 0.0001,
    'learning_rate_critic': 0.001,
    'gamma': 0.95,
    'tau': 0.001,
    'hidden_nodes': [500, 500],
    'batch_norm': False,
    'batch_size': 32,
    'buffer_size': 10000,
    'num_updates_iter': 1
}


    

In [4]:

    

sess = tf.InteractiveSession()


    

In [5]:

    

stat = Statistics(sess, ENV_NAME, ALGO_NAME, DDPG.get_summary_tags(), settings=SETTINGS, save=SAVE)


    

    




/home/bartkeulen/repositories/drl/drl/../results
For visualizing run:
  tensorboard --logdir=/home/bartkeulen/repositories/drl/results/test/TwoLinkArm/DDPG/buffer_size=10000/batch_size=32/batch_norm=False/gamma=0.95/learning_rate_actor=0.0001/learning_rate_critic=0.001/hidden_nodes=[500, 500]/num_updates_iter=1/tau=0.001/0

In [ ]:

    

with tf.Session() as sess:
    env = TwoLinkArm(g=0.)

    stat = Statistics(sess, ENV_NAME, ALGO_NAME, DDPG.get_summary_tags(), settings=SETTINGS, save=SAVE)

#     noise = OrnSteinUhlenbeckNoise(
#         action_dim=env.action_dim,
#         mu=0.,
#         theta=0.05,
#         sigma=0.05)
    noise = WhiteNoise(env.action_dim, 0., 0.1)
    noise = LinearDecay(noise, 100, 125)

    ddpg = DDPG(sess=sess,
                env=env,
                stat=stat,
                exploration=noise,
                **SETTINGS)

    ddpg.train(num_episodes=10000,
               max_steps=200,
               render_env=False)

    sess.close()