OpenAI gym——一款开发和比较RL算法的工具包

1# python 2.7, 复制下面
2$ pip install gym
3
4# python 3.5, 复制下面
5$ pip3 install gym

1# python 2.7, 复制下面
2$ pip install gym[all]
3
4# python 3.5, 复制下面
5$ pip3 install gym[all]

 1import gym
 2from RL_brain import DeepQNetwork
 3
 4env = gym.make('CartPole-v0') #定义使用gym库中的哪一个环境
 5env = env.unwrapped #还原env的原始设置，env外包了一层防作弊层
 6
 7print(env.action_space) #查看这个环境可用的action有多少个
 8print(env.observation_space) #查看这个环境中可用的state的observation有多少个
 9print(env.observation_space.high) #查看observation最高取值
10print(env.observation_space.low) #查看observation最低取值
11
12RL = DeepQNetwork(n_actions=env.action_space.n,
13                  n_features=env.observation_space.shape[0],
14                  learning_rate=0.01, e_greedy=0.9,
15                  replace_target_iter=100, memory_size=2000,
16                  e_greedy_increment=0.001,)
17
18total_steps = 0
19
20
21for i_episode in range(100):
22    #获取回合i_episode第一个observation
23    observation = env.reset()
24    ep_r = 0
25    while True:
26        env.render()#刷新环境
27
28        action = RL.choose_action(observation)#选行为
29
30        observation_, reward, done, info = env.step(action)#获取下一个state
31
32        # x是车的水平位移，所以r1是车越偏离中心，得分（reward）越少
33        # theta是棒子离垂直的角度，角度越大，越不垂直。所以r2是棒越垂直，分(reward)越高
34        # 总reward是r1和r2的结合，既考虑位置也考虑角度，这样DQN学习更有效率
35        x, x_dot, theta, theta_dot = observation_
36
37        r1 = (env.x_threshold - abs(x))/env.x_threshold - 0.8
38        r2 = (env.theta_threshold_radians - abs(theta))/env.theta_threshold_radians - 0.5
39        reward = r1 + r2
40
41        #保存这一组记忆
42        RL.store_transition(observation, action, reward, observation_)
43
44        ep_r += reward
45        if total_steps > 1000:
46            RL.learn()
47
48        if done:
49            print('episode: ', i_episode,
50                  'ep_r: ', round(ep_r, 2),
51                  ' epsilon: ', round(RL.epsilon, 2))
52            break
53
54        observation = observation_
55        total_steps += 1
56
57RL.plot_cost()

 1import gym
 2from RL_brain import DeepQNetwork
 3
 4env = gym.make('MountainCar-v0')
 5env = env.unwrapped
 6
 7print(env.action_space)
 8print(env.observation_space)
 9print(env.observation_space.high)
10print(env.observation_space.low)
11
12RL = DeepQNetwork(n_actions=3, n_features=2, learning_rate=0.001, e_greedy=0.9,
13                  replace_target_iter=300, memory_size=3000,
14                  e_greedy_increment=0.0002,)
15
16total_steps = 0
17
18
19for i_episode in range(10):
20
21    observation = env.reset()
22    ep_r = 0
23    while True:
24        env.render()
25
26        action = RL.choose_action(observation)
27
28        observation_, reward, done, info = env.step(action)
29
30        position, velocity = observation_
31
32        # 车开的越高reward越大
33        reward = abs(position - (-0.5))     # r in [0, 1]
34
35        RL.store_transition(observation, action, reward, observation_)
36
37        if total_steps > 1000:
38            RL.learn()
39
40        ep_r += reward
41        if done:
42            get = '| Get' if observation_[0] >= env.unwrapped.goal_position else '| ----'
43            print('Epi: ', i_episode,
44                  get,
45                  '| Ep_r: ', round(ep_r, 4),
46                  '| Epsilon: ', round(RL.epsilon, 2))
47            break
48
49        observation = observation_
50        total_steps += 1
51
52RL.plot_cost()