跟着开源项目学因果推断——whynot（十四）

素颜马尾好姑娘i 2022-09-14 02:27 127阅读 0赞

### 文章目录 ###

![在这里插入图片描述][2978e569a3424aaf9a0462e0bdaab667.png]

--------------------

WhyNot是一个Python包，它提供了一个用于动态决策的实验沙箱，将因果推理和强化学习工具与具有挑战性的动态环境连接起来。  
该软件包有助于开发、测试、基准测试和教学因果推理和顺序决策工具。

github: https://github.com/zykls/whynot  
文档：https://whynot.readthedocs.io/en/latest/

下载：

pip install whynot

举一个例子 [hiv\_simulator.ipynb][hiv_simulator.ipynb]

%load_ext autoreload
    %autoreload 2
    
    import whynot.gym as gym
    import numpy as np
    import matplotlib.pyplot as plt
    import torch
    
    from scripts import utils
    %matplotlib inline
    
    # HIV环境
    # Make the HIV environment and set random seed.
    env = gym.make('HIV-v0')
    np.random.seed(1)
    env.seed(1)
    torch.manual_seed(1)
    
    # 策略
    class NoTreatmentPolicy(utils.Policy):
        """The policy of always no treatment."""
        def __init__(self):
            super(NoTreatmentPolicy, self).__init__(env)
        
        def sample_action(self, obs):
            return 0
        
    class MaxTreatmentPolicy(utils.Policy):
        """The policy of always applying both RT inhibitor and protease inhibitor."""
        def __init__(self):
            super(MaxTreatmentPolicy, self).__init__(env)
        
        def sample_action(self, obs):
            return 3
        
    class RandomPolicy(utils.Policy):
        """The policy of picking a random action at each time step."""
        def __init__(self):
            super(RandomPolicy, self).__init__(env)
        
        def sample_action(self, obs):
            return np.random.randint(4)
    
    # 模拟
    learned_policy = utils.run_training_loop(
        env=env, n_iter=300, max_episode_length=100, batch_size=1000, learning_rate=1e-3)

几种策略，我们定义了一个基本的Policy类。每个策略都有一个sample\_action函数，它接受一个观察并返回一个操作，一共四类：

*  NNPolicy  
    一种1层前馈神经网络，以状态维为输入维，一个包含8个神经元的隐藏层(状态维度为6)，以动作作为输出维度。我们使用批处理归一化和ReLU激活。
 *  没有任何刺激策略  
    任何调整都不要，属于最基本的
 *  刺激最大策略  
    任何调整都需要，属于过分敏感
 *  随机的政策  
    采取随机行动而不考虑观察结果。  
    ![在这里插入图片描述][278864a77f754ca790065d28bd199427.png]

# 其他三个策略
    policies = {
        "learned_policy": learned_policy,
        "no_treatment": NoTreatmentPolicy(),
        "max_treatment": MaxTreatmentPolicy(),
        "random": RandomPolicy(),
    }
    utils.plot_sample_trajectory(env, policies, 100, wn.hiv.State.variable_names())

![在这里插入图片描述][aaa489803559480e8142a0c6e8fac331.png]

[2978e569a3424aaf9a0462e0bdaab667.png]: /images/20220828/13146c17853741e3a7144e659c4388df.png
[hiv_simulator.ipynb]: https://github.com/zykls/whynot/blob/master/examples/reinforcement_learning/hiv_simulator.ipynb
[278864a77f754ca790065d28bd199427.png]: /images/20220828/e65493b51691423bafc6866d204373da.png
[aaa489803559480e8142a0c6e8fac331.png]: /images/20220828/3c408143ed234d0f8de9c8f47a117d89.png