envs.py 文件源码

def _step(self, action):
        obs, reward, done, info = self.env.step(action)
        obs = self.process_observation(obs)

        if self.squash_rewards:
            reward = float(np.sign(reward))
        else:
            reward = float(reward) / float(self.reward_scale)

        info["frame/lives"] = info["ale.lives"]
        if self.lives is None:
            self.lives = info["ale.lives"]
        else:
            current_lives = info["ale.lives"]
            lost = self.lives - current_lives
            self.lives = current_lives
            if lost > 0:
                reward -= lost * self.death_penalty

        return obs, reward, done, info