Function run

examples/python/learning_tensorflow.py:178–222

(agent, game, replay_memory)

def run(agent, game, replay_memory):
    time_start = time()

    for episode in range(num_train_epochs):
        train_scores = []
        print("\nEpoch %d\n-------" % (episode + 1))

        game.new_episode()

        for i in trange(learning_steps_per_epoch, leave=False):
            state = game.get_state()
            screen_buf = preprocess(state.screen_buffer)
            action = agent.choose_action(screen_buf)
            reward = game.make_action(actions[action], frames_per_action)
            done = game.is_episode_finished()

            if not done:
                next_screen_buf = preprocess(game.get_state().screen_buffer)
            else:
                next_screen_buf = tf.zeros(shape=screen_buf.shape)

            if done:
                train_scores.append(game.get_total_reward())

                game.new_episode()

            replay_memory.append((screen_buf, action, reward, next_screen_buf, done))

            if i >= batch_size:
                agent.train_dqn(get_samples(replay_memory))

            if (i % target_net_update_steps) == 0:
                agent.update_target_net()

        train_scores = np.array(train_scores)
        print(
            "Results: mean: {:.1f}±{:.1f},".format(
                train_scores.mean(), train_scores.std()
            ),
            "min: %.1f," % train_scores.min(),
            "max: %.1f," % train_scores.max(),
        )

        test(test_episodes_per_epoch, game, agent)
        print("Total elapsed time: %.2f minutes" % ((time() - time_start) / 60.0))


def test(test_episodes_per_epoch, game, agent):