Method run

rl2/a3c/worker.py:147–169

(self, sess, coord, t_max)

    self.returns_list = returns_list # Global returns list to plot later

  def run(self, sess, coord, t_max):
    with sess.as_default(), sess.graph.as_default():
      # Assign the initial state
      self.state = repeat_frame(self.img_transformer.transform(self.env.reset()))

      try:
        while not coord.should_stop():
          # Copy weights from  global networks to local networks
          sess.run(self.copy_params_op)

          # Collect some experience
          steps, global_step = self.run_n_steps(t_max, sess)

          # Stop once the max number of global steps has been reached
          if self.max_global_steps is not None and global_step >= self.max_global_steps:
            coord.request_stop()
            return

          # Update the global networks using local gradients
          self.update(steps, sess)

      except tf.errors.CancelledError:
        return

  def sample_action(self, state, sess):
    # Make input N x D (N = 1)