gaussian_gru_policy.py 文件源码

def get_action(self, observation):
        if self._state_include_action:
            if self._prev_action is None:
                prev_action = np.zeros((self.action_space.flat_dim,))
            else:
                prev_action = self.action_space.flatten(self._prev_action)
            all_input = np.concatenate([
                self.observation_space.flatten(observation),
                prev_action
            ])
        else:
            all_input = self.observation_space.flatten(observation)
            # should not be used
            prev_action = np.nan
        mean, log_std, hidden_vec = [x[0] for x in self._f_step_mean_std([all_input], [self._prev_hidden])]
        rnd = np.random.normal(size=mean.shape)
        action = rnd * np.exp(log_std) + mean
        self._prev_action = action
        self._prev_hidden = hidden_vec
        agent_info = dict(mean=mean, log_std=log_std)
        if self._state_include_action:
            agent_info["prev_action"] = prev_action
        return action, agent_info