limited_discrete_action_trainer.py 文件源码

def _update_quantile(self):
        states = np.array(self._quantile_states, dtype=np.float32)
        limited_action_values = self.action_values(states, self._limited_action)
        base_action_values = np.max(
            np.array(
                [
                    self.action_values(states, action)
                    for action in six.moves.range(self.num_actions)
                    if action != self._limited_action
                ]
            ),
            axis=0
        )
        target = np.percentile(
            limited_action_values - base_action_values, self._quantile
        )
        print("REWARD PENALTY TARGET:", target)
        self.quantile_value += self._quantile_update_rate * target
        print("QUANTILE:", self.quantile_value)