num_consecutive_iterations = 100
last_time_steps = np.zeros(
num_consecutive_iterations) # 只存储最近100场的得分（可以理解为是一个容量为100的栈）
env = gym.make('CartPole-v0')
for episode in range(5000):
observation = env.reset() # 初始化本场游戏的环境
episode_reward = 0
action = q_f.get_actions(observation)
next_action = action
for t in range(max_number_of_steps):
# env.render()
next_observation, reward, done, info = env.step(action)
next_action = q_f.get_actions(next_observation)
action = q_f.learn(observation, action, next_observation, next_action,
-200 if done else reward)
observation = next_observation
action = next_action
episode_reward += reward
if done:
print('%d Episode finished after %f time steps / mean %f' %
(episode, t + 1, last_time_steps.mean()))
last_time_steps = np.hstack(
(last_time_steps[1:], [episode_reward])) # 更新最近100场游戏的得分stack
break
# 如果最近100场平均得分高于195
if (last_time_steps.mean() >= goal_average_steps):
print('Episode %d train agent successfuly!' % episode)
break
learning_rate=0.8,
discount_rate=0.9,
random_action_prob=0.5,
random_action_decay_rate=0.99,
dyna_iterations=0)
start_state = gw.grid_coordinates_to_indices(start)
iterations = 1000
### IMPORTANT
# you do need to write your own generate_experience function
# based on either \epilon greedy or exploration function
# make sure on your submission, you need to submit
# a new rl_qlearn.py that has the updated gw.generate_experience_your_name
### IMPORTANT
flat_policies, flat_utilities = sa.learn(start_state,
gw.generate_experience,
iterations=iterations)
new_shape = (gw.shape[0], gw.shape[1], iterations)
sa_utility_grids = flat_utilities.reshape(new_shape)
sa_policy_grids = flat_policies.reshape(new_shape)
print('Final result of Sarsa:')
print(sa_policy_grids[:, :, -1])
print(sa_utility_grids[:, :, -1])
plt.figure()
gw.plot_policy(sa_utility_grids[:, :, -1], sa_policy_grids[:, :, -1])
plot_convergence(sa_utility_grids, sa_policy_grids)
plt.show()