Method gradient

examples/reinforcement_learning/tutorial_TRPO.py:253–267

pi gradients :param states: state batch :param actions: actions batch :param adv: advantage batch :param old_log_prob: old log probability batch :return: gradient

(self, states, actions, adv, old_log_prob)

        return -surr

    def gradient(self, states, actions, adv, old_log_prob):
        """
        pi gradients
        :param states: state batch
        :param actions: actions batch
        :param adv: advantage batch
        :param old_log_prob: old log probability batch
        :return: gradient
        """
        pi_params = self.actor.trainable_weights
        with tf.GradientTape() as tape:
            loss = self.pi_loss(states, actions, adv, old_log_prob)
        grad = tape.gradient(loss, pi_params)
        gradient = self._flat_concat(grad)
        return gradient, loss

    def train_vf(self, states, rewards_to_go):
        """