acer.py 文件源码

def __call__(self, obs):
        action_distrib = self.pi(obs)
        v = self.v(obs)

        def evaluator(action):
            adv_mean = sum(self.adv(obs, action_distrib.sample().data)
                           for _ in range(self.n)) / self.n
            return v + self.adv(obs, action) - adv_mean

        action_value = SingleActionValue(evaluator)

        return action_distrib, action_value, v