fbf3bafd970d3539ad774bae5f46a4b730901cb5,tensorforce/updater/trpo_updater.py,TRPOUpdater,create_training_operations,#TRPOUpdater#,99

Before Change

            mean_kl_divergence = get_kl_divergence_gaussian(self.prev_action_means, self.prev_action_log_stds,
                                                            self.action_means, self.action_log_stds) / float(
                self.batch_size)
            mean_entropy = get_entropy_gaussian(self.action_log_stds) / float(self.batch_size)

            self.losses = [surrogate_loss, mean_kl_divergence, mean_entropy]

            // Get symbolic gradient expressions

After Change

            surrogate_loss = -tf.reduce_mean(prob_ratio * self.advantage)
            variables = tf.trainable_variables()
            batch_float = tf.cast(self.batch_size, tf.float32)

            mean_kl_divergence = get_kl_divergence_gaussian(self.prev_action_means, self.prev_action_log_stds,
                                                            self.action_means, self.action_log_stds) / batch_float
            mean_entropy = get_entropy_gaussian(self.action_log_stds) / batch_float

Instances