Method run_n_steps

rl2/a3c/worker.py:185–217

(self, n, sess)

    return vhat[0]

  def run_n_steps(self, n, sess):
    steps = []
    for _ in range(n):
      # Take a step
      action = self.sample_action(self.state, sess)
      next_frame, reward, done, _ = self.env.step(action)

      # Shift the state to include the latest frame
      next_state = shift_frames(self.state, self.img_transformer.transform(next_frame))

      # Save total return
      if done:
        print("Total reward:", self.total_reward, "Worker:", self.name)
        self.returns_list.append(self.total_reward)
        if len(self.returns_list) > 0 and len(self.returns_list) % 100 == 0:
          print("*** Total average reward (last 100):", np.mean(self.returns_list[-100:]), "Collected so far:", len(self.returns_list))
        self.total_reward = 0.
      else:
        self.total_reward += reward

      # Save step
      step = Step(self.state, action, reward, next_state, done)
      steps.append(step)

      # Increase local and global counters
      global_step = next(self.global_counter)

      if done:
        self.state = repeat_frame(self.img_transformer.transform(self.env.reset()))
        break
      else:
        self.state = next_state
    return steps, global_step

  def update(self, steps, sess):
    """