Once the agent has been trained, we evaluate its performance over 100 new episodes. This can be done by calling the .test function and feeding the arguments environment and number of episodes on which to test.

# Evaluate the agent on 100 new episodes.
scores = sarsa.test(env, nb_episodes=100, visualize=False)

print('Average score over 100 test games: {}'.format(np.mean(scores.history['episode_reward'])))