Reinforcement Learning with Pytorch. [Error: KeyError ]

Question

我是强化学习和pytorch的新手。我正在从 Udemy 学习它。但是，我拥有的代码与显示的代码相同，但出现错误。我猜这是一个 pytorch 错误，但无法调试它。如果有人提供帮助，我们将不胜感激。

import gym
import time
import torch
import matplotlib.pyplot as plt
from gym.envs.registration import register
register(id='FrozenLakeNotSlippery-v0',entry_point='gym.envs.toy_text:FrozenLakeEnv',kwargs={'map_name': '4x4', 'is_slippery':False})
env = gym.make('FrozenLakeNotSlippery-v0')
number_of_states = env.observation_space.n
number_of_actions = env.action_space.n
Q = torch.zeros([number_of_states,number_of_actions])
num_episodes = 1000
steps_total = []
gamma = 1
for i in range(num_episodes):
    state = env.reset()
    step = 0
    while True:
        step += 1
        #action = env.action_space.sample()
        random_values = Q[state]+torch.rand(1,number_of_actions)/1000
        action = torch.max(random_values,1)[1][0]
        new_state, reward, done, info = env.step(action)
        Q[state, action] = reward + gamma * torch.max(Q[new_state])
        state = new_state
        #time.sleep(0.4)
        #env.render()
        if done:
            steps_total.append(step)
            print ("Episode Finished after %i steps" %step)
            break

print ("Average Num Steps: %2f" %(sum(steps_total)/num_episodes))
plt.plot(steps_total)
plt.show()

我遇到的错误是以下错误

KeyError                                  Traceback (most recent call last)
<ipython-input-11-a6aa419c3767> in <module>
  8         random_values = Q[state]+torch.rand(1,number_of_actions)/1000
  9         action = torch.max(random_values,1)[1][0]
---> 10         new_state, reward, done, info = env.step(action)
 11         Q[state, action] = reward + gamma * torch.max(Q[new_state])
 12         state = new_state

c:\users\souradip\appdata\local\programs\python\python36\lib\site-packages\gym\envs\toy_text\discrete.py in step(self, a)
 53 
 54     def step(self, a):
---> 55         transitions = self.P[self.s][a]
 56         i = categorical_sample([t[0] for t in transitions], self.np_random)
 57         p, s, r, d= transitions[i]

KeyError: tensor(3)

Answer 1

下面的代码

action = torch.max(random_values,1)[1][0]

得到一个 0-dim 张量，但是 env.step() 需要一个 python 数字，这基本上是来自动作 space 的一个动作。因此，正如评论中提到的 @a_guest，使用 a.item() 将 0-dim 张量转换为 python 数字，如下所示：

new_state, reward, done, info = env.step(action.item())

Reinforcement Learning with Pytorch. [Error: KeyError ]

Reinforcement Learning with Pytorch. [Error: KeyError ]

reinforcement-learning

python-3.x

pytorch