a95eb8bd878b43346c4bb2e8e29911dc1ab90638,A2C.py,A2CContinuous,learn,#A2CContinuous#,201

Before Change

            returns = np.concatenate([discount_rewards(trajectory["reward"], config["gamma"]) for trajectory in trajectories])
            qw_new = self.get_critic_value(all_state)

            self.sess.run([self.critic_train], feed_dict={self.critic_state_in: all_state, self.critic_target: returns})
            target = np.mean((returns - qw_new) ** 2)
            self.sess.run([self.actor_train], feed_dict={self.input_state: all_state, self.actions_taken: all_action, self.target: target})

            episode_rewards = np.array([trajectory["reward"].sum() for trajectory in trajectories])  // episode total rewards

After Change

            episode_rewards = np.array([trajectory["reward"].sum() for trajectory in trajectories])  // episode total rewards
            episode_lengths = np.array([len(trajectory["reward"]) for trajectory in trajectories])  // episode lengths

            results = self.sess.run([self.summary_op, self.critic_train, self.actor_train], feed_dict={
                                    self.critic_state_in: all_state,
                                    self.critic_target: returns,
                                    self.input_state: all_state,
                                    self.actions_taken: all_action,
                                    self.critic_feedback: qw_new,
                                    self.critic_rewards: returns,
                                    self.rewards: np.mean(episode_rewards),
                                    self.episode_lengths: np.mean(episode_lengths)
                                    })
            self.writer.add_summary(results[0], iteration)
            self.writer.flush()

            reporter.print_iteration_stats(iteration, episode_rewards, episode_lengths, total_n_trajectories)

Instances