强化学习实战（一）—— 使用BaslineDQN学习飞船降落

本文将介绍如何使用Stable Basline3中的DQN算法学习飞船降落问题。

实验过程1

1. 引入库并创建环境

import gym
from stable_baselines3 import DQN
from stable_baselines3.common.evaluation import evaluate_policy

# Create environment
env = gym.make("LunarLander-v2")
1
2
3
4
5
6

（1）了解环境信息
成功创建环境后，我们可以通过env.action_space和env.observation_space查看环境的动作和状态空间。也可以通过env.action_space.sample()和env.observation_space.sample()随机采样，了解具体的动作和状态表示情况。

print(env.action_space)
print(env.action_space.sample())

print(env.observation_space)
print(env.observation_space.sample())
1
2
3
4
5

Discrete(4)
3
Box([-inf -inf -inf -inf -inf -inf -inf -inf], [inf inf inf inf inf inf inf inf], (8,), float32)
[-0.39453888  0.88357323 -2.6758633   0.26985604 -0.31590447 -0.5141233
  1.2682225   0.7396759 ]
1
2
3
4
5

接着我们可以通过查询gym官方文档了解动作和状态空间更细节的信息。

2. 创建模型

# Instantiate the agent
model = DQN("MlpPolicy",
            env,
            tensorboard_log = './logs',
            verbose=1)
1
2
3
4
5

Parameters
下面将对一些特殊的常用的参数进行说明：

• policy – The policy model to use
  ①MlpPolicy：DQNPolicy
  ②CnnPolicy：Policy class for DQN when using images as input.
  ③MultiInputPolicy：Policy class for DQN when using dict observations as input.

• tensorboard_log – the log location for tensorboard (if None, no logging)

• verbose – Verbosity level: 0 for no output, 1 for info messages (such as device or wrappers used), 2 for debug messages

• policy_kwargs (Optional[Dict[str, Any]]) – additional arguments to be passed to the policy on creation

• seed – Seed for the pseudo random generators

• device – Device (cpu, cuda, …) on which the code should be run. Setting it to auto, the code will be run on the GPU if possible.

3. 模型学习

# Train the agent and display a progress bar
model.learn(total_timesteps=int(5e5),
            tb_log_name = 'DQN2',
            progress_bar= True,)
# Save the agent
model.save("dqn_lunar")
del model  # delete trained model to demonstrate loading
1
2
3
4
5
6
7

Parameters
下面将对一些特殊的常用的参数进行说明：

• total_timesteps – The total number of samples (env steps) to train on
• log_interval (int) – The number of timesteps before logging.
• tb_log_name (str) – the name of the run for TensorBoard logging
• progress_bar (bool) – Display a progress bar using tqdm and rich.

4. 模型评估

方式一

# Load the trained agent
# model = DQN.load("dqn_lunar", env=env, print_system_info=True)
model = DQN.load("dqn_lunar", env=env)
# Evaluate the agent
# NOTE: If you use wrappers with your environment that modify rewards,
#       this will be reflected here. To evaluate with original rewards,
#       wrap environment in a "Monitor" wrapper before other wrappers.
mean_reward, std_reward = evaluate_policy(model, 
                                          model.get_env(), 
                                          render = True,
                                          n_eval_episodes=10)
1
2
3
4
5
6
7
8
9
10
11

方式二

model = DQN.load("dqn_lunar", env=env)
# Evaluate the agent
episodes = 10
for ep in range(episodes):
    obs = env.reset()
    done = False
    rewards = 0
    while not done:
#         action = env.action_space.sample()
        action, _states = model.predict(obs, deterministic=True)
        obs,reward,done,info = env.step(action)
        env.render()
        rewards += reward
    print(rewards)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

5. 附录

Code

import gym
from stable_baselines3 import DQN
from stable_baselines3.common.evaluation import evaluate_policy

# Create environment
env = gym.make("LunarLander-v2")
print(env.action_space)
print(env.action_space.sample())
# do nothing, fire left orientation engine, fire main engine, fire right orientation engine.
print(env.observation_space)
print(env.observation_space.sample())
# the coordinates of the lander in x & y, its linear velocities in x & y, its angle, its angular velocity, 
# and two booleans that represent whether each leg is in contact with the ground or not.
# Instantiate the agent
model = DQN("MlpPolicy",
            env,
            tensorboard_log = './logs',
            verbose=1)
# Train the agent and display a progress bar
model.learn(total_timesteps=int(5e5),
            tb_log_name = 'DQN2',
            progress_bar= True,)
# Save the agent
model.save("dqn_lunar")
del model  # delete trained model to demonstrate loading

# Load the trained agent
# NOTE: if you have loading issue, you can pass `print_system_info=True`
# to compare the system on which the model was trained vs the current one
# model = DQN.load("dqn_lunar", env=env, print_system_info=True)
model = DQN.load("dqn_lunar", env=env)
# Evaluate the agent
episodes = 10
for ep in range(episodes):
    obs = env.reset()
    done = False
    rewards = 0
    while not done:
#         action = env.action_space.sample()
        action, _states = model.predict(obs, deterministic=True)
        obs,reward,done,info = env.step(action)
#         env.render()
        rewards += reward
    print(rewards)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45

Result

很明显飞船表现得并不好，当它下降到一定位置后便开始悬浮，不符合要求。我们需要修改训练参数。

实验过程2

1. 修改模型参数

（1）加大学习率learning_rate，加快收敛。
（2）修改网络结构
可以通过policy_kwargs传递网络参数，通过查看MLPPolicy参数可知net_arch可以修改网络结构。

网络结构修改参考

查看源码我们知道初始网络结构为[64, 64]，因此我们修改结构为[256, 256]。

# Instantiate the agent
model = DQN("MlpPolicy",
            env,
            tensorboard_log = './logs',
            learning_rate = 5e-4,
            policy_kwargs = {'net_arch': [256,256]},
            verbose=1)
1
2
3
4
5
6
7