ddpg_cartpole.py 文件源码

def __init__(self, namespace, input_state, action_dim):
    super(ActorNetwork, self).__init__(namespace)

    self.input_state = input_state

    self.exploration_noise = util.OrnsteinUhlenbeckNoise(action_dim, 
                                                         opts.action_noise_theta,
                                                         opts.action_noise_sigma)

    with tf.variable_scope(namespace):
      opts.hidden_layers = opts.actor_hidden_layers
      final_hidden = self.input_state_network(self.input_state, opts)
      # action dim output. note: actors out is (-1, 1) and scaled in env as required.
      weights_initializer = tf.random_uniform_initializer(-0.001, 0.001)
      self.output_action = slim.fully_connected(scope='output_action',
                                                inputs=final_hidden,
                                                num_outputs=action_dim,
                                                weights_initializer=weights_initializer,
                                                weights_regularizer=tf.contrib.layers.l2_regularizer(0.01),
                                                activation_fn=tf.nn.tanh)