Wrong A3C implementation about ga3c

[wgeul]

I think that the overall design is of a different approach:

            prediction, value = self.predict(self.env.current_state)
            action = self.select_action(prediction)
            reward, done = self.env.step(action)
            reward_sum += reward
            exp = Experience(self.env.previous_state, action, prediction, reward, done)

The last line indicates that experience is backward looking. I assume that's why the terminal_reward that is equal to value is consistent when done is False?